Chapter 10

Forecasting Demand, Sales, and Campaign Performance

From Customers to Calendars

Dr. Jose Mendoza, Academic Director and Clinical Associate Professor

Version 1.0 · July 2026

Except where otherwise noted, this chapter is licensed under CC BY 4.0.

Chapter Information

ABSTRACT

This chapter develops forecasting as the time-indexed branch of the predictive discipline built in Chapters 8 and 9, with StyleCraft’s holiday open-to-buy plan as its running decision. It establishes the time series as a data structure whose order carries information, uses decomposition to read trend, seasonality, cycle, and noise, and separates recurring calendar effects from structural breaks. It builds the forecasting ladder from the naive and seasonal-naive opponents through the exponential smoothing family, then rebuilds honest evaluation for ordered data: rolling-origin selection, a sealed holdout that resembles the decision, MAPE with its denominator failure, and accuracy reported by horizon. The forecast widens into a distribution through simulated paths and scenarios, and the human hand is disciplined by the written override. The chapter closes with managerial interpretation, industry practice, the ethics of the laundered guess, and exercises.

KEYWORDS

forecasting; time series; seasonality; structural break; seasonal-naive; exponential smoothing; MAPE; prediction interval; forecast horizon; judgmental adjustment

VERSION AND DATE

Version 1.0 · July 2026 · Language: English (United States)

SUGGESTED CITATION

Mendoza, J. (2026). Forecasting demand, sales, and campaign performance. In Applied business analytics for marketing decision-making: Business analytics and data visualization (Chapter 10, Version 1.0) [Open educational resource]. CC BY 4.0.

LICENSE AND RIGHTS

Copyright © 2026 Jose Mendoza. Except where otherwise noted, this work is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0). You may share and adapt this material for any purpose, provided appropriate credit is given. Third-party trademarks, screenshots, figures, and other materials remain subject to their respective rights and licenses.

Google Colab is a product of Google LLC. "Python" and the Python logos are trademarks or registered trademarks of the Python Software Foundation. pandas, NumPy, Matplotlib, and scikit-learn are sponsored or affiliated projects of NumFOCUS, a 501(c)(3) nonprofit charity in the United States. statsmodels is a community-developed project distributed under the modified BSD license. ChatGPT is a product of OpenAI, Claude of Anthropic, Gemini and NotebookLM of Google LLC, and GitHub Copilot of GitHub, Inc. Product names are used for identification only and do not imply endorsement. StyleCraft Collective is a fictional company created for instruction.

COMPANION REPOSITORY

Datasets, notebooks, and figure sources for this chapter: Applied Business Analytics companion repository on GitHub

GENERATIVE AI USE

Generative artificial intelligence and other AI-assisted tools were used in the research, writing, revision, and production of this chapter, including literature discovery, source organization, outlining, preliminary drafts, prose revision, support for code and analytical examples, and document formatting. These tools were used under the author’s direction and are not credited as authors, researchers, or sources. The author determined the chapter’s scope, learning objectives, methods, interpretations, and recommendations, and reviewed and approved all AI-assisted material: factual claims and citations were checked against the underlying sources rather than accepted from AI-generated summaries, and code and analytical outputs were tested or otherwise reviewed for accuracy. Responsibility for the accuracy, originality, and final form of this chapter rests entirely with the author. A fuller statement appears in the front matter of the complete guide.

Chapter Learning Objectives

By the end of this chapter, students should be able to:

• Explain what makes a time series a distinct data structure — ordered observations whose sequence carries information — and construct a daily revenue series from transaction-grain data by a deliberate grain change, per Section 3.3 and the groupby machinery of Chapter 4.
• Name the four components of a marketing time series — trend, seasonality, cycle, and noise — read each from a decomposition plot at concept level, and state which component each forecasting method in this chapter is built to track.
• Distinguish recurring calendar effects from structural breaks, explain why history from before a break misleads a model after it, and identify StyleCraft’s store-opening inflection as a designed break announced in the company’s own data.
• State the boundary between the weekly seasonality a period-7 method carries and the annual, moving, and promotional calendar effects it does not, and name the three instruments that carry the second group.
• Construct naive and seasonal-naive forecasts as the time extension of the baselines of Section 8.7 and defend the seasonal-naive rule as the opponent every forecasting model must beat before it earns a planning decision.
• Describe the exponential smoothing family at working level — simple smoothing, Holt’s trend method, damped trend, and Holt-Winters seasonal smoothing — and state what each added layer tracks, what it costs, and how each can overfit.
• Explain why random train/test splits fail for ordered data and build the two-layer evaluation this chapter requires: rolling-origin comparison inside the training data for method selection, and one sealed final holdout, chosen to resemble the decision, for the grade.
• Compute and interpret forecast accuracy with MAPE alongside the MAE and RMSE of Section 8.6, guard MAPE’s denominator in code, and report accuracy as a function of horizon estimated across several forecast origins rather than one.
• Distinguish a point forecast from the distribution it summarizes, build daily, weekly, and season-total prediction intervals by aggregating simulated paths before taking quantiles, and state what a simulation band does and does not price.
• State when a judgmental adjustment to a statistical forecast is legitimate, size an override as a ramp from controlled evidence with a stated range, and apply the six-element written-override discipline.
• Recognize the three standing forecasting failure modes — extrapolating through structural breaks, overfitting seasonality, and ignoring promotions — and diagnose each from a forecast exhibit.
• Verify an AI-drafted forecasting pipeline with the five-point audit of Section 8.11 plus the forecasting supplement, predicting the holdout error before running.
• Assemble a forecast deliverable a planning meeting can hold accountable: method and opponent stated that sealed-holdout grade in MAPE and dollars, intervals communicated in words, overrides documented, and the forecast translated into a declared planning unit whose limits are stated.

Chapter 9 ended with a scored file and a signature. The Comeback Edit’s targeting rule was locked — the churn label defined in writing, the threshold re-derived from the program’s own twelve-dollar economics, the fairness appendix catching the model before it condemned the newest stores for being new — and the analyst of record signed a deliverable that made predictions about individual people and owned what the program would do to them. Every prediction in that deliverable shared a quiet structural assumption: the customers were exchangeable. The split shuffled them freely, the model scored them in any order, and nothing about customer C004471’s row depended on which row came before it. This chapter forfeits the assumption, because the next family of marketing questions abandons the individual for the aggregate and the snapshot for the flow of time itself. What will revenue be in December? How will the new stores’ ramp reshape the season? What should the buy plan assume about demand that has not happened yet? The observations are no longer eight thousand interchangeable customers but roughly seven hundred and thirty ordered days, each one glued to its neighbors by trend, rhythm, and event — and the ordering is not an inconvenience to be shuffled away but the single most informative thing about the data. The machinery of Part II rides along where it can: the frame is still declared before fitting, the baseline is still named before the model exists, the test data is still the future the model has not seen. But time rewrites the mechanics of every one of those disciplines, adds two instruments the customer-grain chapters never needed — the decomposition that separates a series into its stories, and the prediction interval that keeps a forecast honest — and installs this chapter’s hardest lesson at the exact place StyleCraft’s own history planted it: a model trained on the past can only extend the past, and when the business has deliberately broken with its past — an expansion, an opening, a new market — the break is something the analyst must bring to the model, because no amount of fitting will make the model see it on its own.

concept

What This Chapter Is Really About

Chapters 8 and 9 graded models against a future made of held-out customers — a future the analyst could manufacture at will, because customers are exchangeable and any random twenty percent of them stands in for the rest. Time offers no such courtesy. The future of a time series is not a random sample of its past; it is the next segment of a single unrepeatable path, connected to everything before it and different from all of it. That one fact rearranges the chapter. Evaluation must move the wall of Section 8.3 from between customers to between yesterday and tomorrow, because a shuffled split lets the model interpolate — predict a missing Tuesday from the Tuesdays on either side of it — which is a party trick, not a forecast. The baseline must learn the calendar, because “tomorrow resembles today” is free, obvious, and brutally hard to beat. And the model must be caught at the boundary of its competence: every method in this chapter works by extending patterns the history contains, which means every method is structurally blind to the two things StyleCraft’s planning team most needs it to see — the deliberate break with history that the expansion is, and the annual holiday shape that two Decembers cannot teach it.

The chapter’s real subject is the resulting division of labor. The model owns the historical rhymes: level, trend, and the weekly rhythm. The analyst owns the future calendar the model cannot read: the announced opening, the moving holiday, the planned promotion, the future the company intends. And the forecast the meeting can trust is the one where both contributions are visible, graded, and signed — the model’s on a holdout chosen to resemble the decision, the analyst’s in a written override.

Source: Course concept developed for this guide, informed by Hyndman and Athanasopoulos (2021) and Fildes et al. (2009).

10.1 Marketing Decision Context: The Buy Plan That Assumes a Future

Every deliverable in Part II so far has predicted people. The Backstage list ranked customers by expected spend; the Comeback Edit ranked them by risk of leaving. The decision that opens this chapter predicts no one in particular — and commits more money than both programs combined. It is late summer at StyleCraft, and the head of merchandise planning owns the two documents that will define the company’s fourth quarter. The first is the holiday open-to-buy plan: the dollar envelope of receipts the November–December assortment may commit, which must be signed roughly six weeks from now because apparel lead times are long and the factories’ calendars do not negotiate. The second is the Q4 media pacing plan: how the season’s marketing budget spreads across the weeks from late October through the December peak. Both documents assume a future. An open-to-buy envelope is a bet on December demand placed in August; a pacing plan is a bet on which weeks the demand will arrive in. And the bet is bigger and stranger this year than it has ever been, because the number both plans need — expected revenue, week by week, through the holiday season — must span the thing StyleCraft has spent two years doing to itself: the expansion. Eight Wave-4 stores have opened across the trailing year, several of them recently enough that their ramp curves are still visibly climbing in the weekly numbers. Whatever last December looked like, this December will not look like it, by design.

Three forecasts are already circulating, and — the pattern is by now the guide’s oldest running joke — each arrives confident and none is accountable. The first is finance’s trendline: last year’s holiday revenue, grown by the year-over-year rate of the spring, drawn forward as a straight line. It is hopeful in the precise sense that it assumes the future will be the past plus a constant, and it makes no statement about what would have to be true for that to hold. The second lives inside the planning module of StyleCraft’s inventory software, which offers a Demand Forecast button and, when pressed, produces a number — a single number, to the dollar, with no interval, no stated method, and documentation that says only that it uses “advanced machine learning.” The third belongs to the merchant who has been right before: the founder’s longtime head of buying, whose position is that the suburban stores will “come alive for holiday gifting” and that the season will be up thirty percent. She may well be right. But her number cannot be interrogated, cannot be graded, and — the planning lead’s actual complaint — cannot be defended to the CFO when the receipts it justified become markdowns in January. Three numbers, three provenances, and not one of them can answer the only question that matters about a forecast before you bet on it: how wrong does this method tend to be, and how would we know?

So, the planning lead has borrowed the VP of Marketing’s analyst — still you — and commissioned the deliverable this chapter builds. Not a number: a forecast, in this guide’s full sense. A stated method, fitted to the daily revenue history that StyleCraft’s own transactions generate, and chosen by comparison rather than preference. A stated opponent — the seasonal-naive rule this chapter will introduce, which any funded model must visibly beat. A grade earned the only way a seasonal forecast can be graded honestly: by holding out a stretch of the known past that resembles the decision, forecasting it as if it were still unknown, and measuring the miss — including the test the expansion makes non-negotiable, whether the method sees the inflection the store openings printed into the series, and what it takes from the analyst to make it see the next one. Prediction intervals, because the plan needs to know not just the center of the future but its plausible width, and because a point forecast delivered without one is — Section 10.15 will use the hard work — a laundered guess. And a written override log, because the analyst knows things the history cannot: the openings already announced for the fall, the promotion calendar marketing has drafted, the moving date of Thanksgiving.

One boundary belongs in the brief rather than in a footnote nine sections later, because it constrains what the whole project can claim. This chapter forecasts chain-level revenue. A revenue forecast can size an open-to-buy envelope in dollars and pace media by week. It cannot, on its own, tell a buyer how many units of which dress in which size to send to which store, because that translation needs realized price, category and size mix, margin, markdown expectation, opening inventory, and store allocation — none of which a chain-level revenue series contains. The program is therefore commissioned as a revenue-and-pacing deliverable with a declared dollar envelope, and the unit-level assortment buy is named as a separate forecast at a separate grain, with its own baselines and its own holdout. Naming the boundary in the brief is what keeps Section 10.13 from promising quantities the analysis cannot produce.

The receipts lock in six weeks. Sections 10.2 through 10.4 teach the reader to see the series. Sections 10.5 and 10.6 build the forecasting ladder from free baselines to Holt-Winters. Sections 10.7 and 10.8 rebuild honest evaluation for ordered data and widen the forecast from a number into a distribution. Section 10.9 disciplines the human hand. Section 10.10 catalogs the failures. Section 10.11 turns to the AI assistant, which will draft any of it fluently and grade none of it honestly. The labs in Section 10.12 run the whole discipline on the certified series, and Sections 10.13 through 10.15 rehearse the meeting where a forecast becomes a commitment — and someone becomes accountable for the difference.

10.2 Time Series as a Data Structure

The opening case asked for a forecast of revenue by week; this section starts further back, with the shape of the data the forecast will stand on, because the shape is the chapter’s first genuinely new idea. Every dataset in Part II so far has been a table of exchangeable rows. The eight thousand customers of the feature table could be shuffled without destroying anything: no customer’s row depended on which row preceded it, and the train/test split of Section 8.5 exploited exactly that freedom. A time series abandons the freedom. Its observations are indexed by time — for StyleCraft, one row per calendar day — and the index is not a label but a structure: each day’s revenue is stitched to its neighbors by the trend the business is riding, the weekly rhythm of retail traffic, the season the calendar imposes, and the aftermath of whatever happened yesterday. Shuffle the rows of a time series and the object is destroyed; the information was in the order.

definition

Time Series

A time series is a sequence of observations of the same quantity, recorded at successive points in time — usually at a regular interval such as a day, week, or month — in which the temporal ordering is part of the data’s meaning. The order must be preserved because observations may exhibit serial dependence, trend, seasonality, or structural change, and because those are the patterns a forecast extends. Not every time series is serially dependent — a series of independent draws recorded in time order is still a time series — but the business series this chapter forecasts almost always are, and any procedure that treats their observations as exchangeable discards the temporal context needed to diagnose and forecast them.

Source: Adapted from Hyndman and Athanasopoulos (2021).

In other words, time order is information — and that is the sentence that separates this chapter from the two before it. In the customer-grain chapters, knowing that one row was “next to” another told you nothing; here it tells you nearly everything, because tomorrow is constrained by today in a way that one customer never constrained another. The constraint is what a forecaster harvests: the whole enterprise of this chapter is the disciplined use of a series’ own past as evidence about its future, and the methods of Sections 10.5 and 10.6 differ mainly in how much of the past they consult and how quickly they forget it.

Before any of that, the series must exist, and for StyleCraft it does not exist natively — it must be manufactured by a grain change. The company’s source of truth is the transactions fact table, one row per order line, per Section 3.3’s grain discipline; a daily revenue series is that table aggregated to one row per day, summing line_revenue within each order date — the deliberate grain change of Section 3.3, executed with the groupby machinery Chapter 4’s labs built and this guide does not re-teach. The certified dataset ships the result as marketing_daily, the daily table that joins the transaction roll-up with campaign activity, and Code 10.3 performs the reconciliation this guide’s known-truth discipline requires: rebuild the daily revenue column from the fact table yourself, and confirm the shipped series matches to the cent, because a forecast inherits every defect of the aggregation underneath it.

Two properties of the manufactured series deserve a sentence each. The grain is a choice, and it is part of the frame: the open-to-buy plan thinks in weeks, the pacing plan in days, and a series can always be aggregated upward (days to weeks) but never disaggregated downward without new assumptions — so this chapter works at the daily grain and rolls up. And the series must be complete: a day with no transactions is a revenue of zero, not a missing row, for exactly the reason a customer with no outcome-window purchases was a label of zero in Section 8.3 — absence is an observation and dropping it would teach the model a calendar with holes in it. That rule has a mechanical consequence the lab enforces rather than assumes reindexing a series onto a complete daily calendar inserts missing days as blanks, not as zeros, and turning the blanks into zeros is a decision the analyst makes and states, not a default the software supplies.

10.3 The Components of a Marketing Time Series

Plot StyleCraft’s daily revenue across its twenty-four months and the line looks, at first, like noise with a mood. The skill this section installs are seeing the line as the sum of separable stories — because every forecasting method in this chapter is, underneath its arithmetic, a claim about which of those stories will continue. The decomposition tradition gives the stories standard names, and this guide states them once, formally, in the box below, then spends the section teaching the reader to point at each one on StyleCraft’s own plot.

definition

Trend, Seasonality, and Cycle

The trend of a time series is its long-run direction — the smooth underlying path the series follows once short-run fluctuation is set aside. Seasonality is a repeating pattern with a fixed and known period: the rhythm of days of the week, or of months of the year, that recurs on schedule at a constant number of observations apart. A cycle is a rise and fall that recurs without a fixed period — economic expansions, fashion waves — distinguishable from seasonality precisely because its timing cannot be read off a calendar. What remains when trend, seasonality, and cycle are accounted for is noise (the remainder or irregular component): the unexplained residual variation no pattern claims.

Source: Adapted from Hyndman and Athanasopoulos (2021).

Each component has a StyleCraft face, and the dataset’s designers planted all four deliberately. The trend is the expansion story itself: the urban business’s steady organic growth, plus the stair-step contributions of each opening — a trend with a complication Section 10.4 will name. The seasonality operates at two nested periods, and the difference between them organizes the rest of the chapter. The weekly rhythm is retail’s heartbeat: quiet early weekdays, a build through Friday, a weekend peak — a pattern that repeats roughly one hundred and four times in the history, which is why every method in this chapter will be given a seasonal period of seven. The annual rhythm is the merchandising calendar: the back-to-school lift and the November–December holiday surge, patterns the history contains only twice — a scarcity with consequences Sections 10.4 and 10.10 will price, and the reason this chapter never asks a period-7 method to carry them. Cycles, at this guide’s concept level, are the honest admission that not every wave has a schedule: a viral product moment, a macroeconomic squeeze on discretionary apparel, a competitor’s collapse. The twenty-four-month window is too short to estimate cycles, and this chapter does not try; the term is owned here so the reader stops mislabeling every unexplained wave “seasonality.” And the noise is everything else — weather, whim, and the irreducible randomness of a few hundred purchase decisions a day — the component no method forecasts and every honest interval respects.

The instrument that makes the components visible is decomposition, and this guide uses it at concept level: as a reading tool, not a forecasting engine. A decomposition procedure takes the observed series and splits it into estimated trend, seasonal, and remainder parts — classical procedures do this with moving averages, and the modern standard, STL, does it with iterated local smoothing (Cleveland et al., 1990) — under an assumed structure, most simply the additive one: observed value equals trend plus seasonal effect plus remainder. The mechanics stay in the software; what the analyst owns is the reading. A decomposition plot stacks four panels — observed, trend, seasonal, remainder — and Code 10.4 will put StyleCraft’s own on screen with the store-opening dates drawn on it. Read the trend panel for the expansion: the climb, and the visible steps where openings land. Read the seasonal panel for the heartbeat: the seven-day sawtooth, regular as a pulse, whose peak-to-trough height in dollars is exactly what the seasonal-naive rule of Section 10.5 carries for free. And read the remainder panel the way Chapter 4 taught you to read every residual exhibit: not as garbage but as a question list, because a spike in the remainder is a day the components could not explain — a promotion the decomposition does not know about, a storm, a data defect — and Section 10.10’s third failure mode begins with an analyst who never looked.

definition

Decomposition

Decomposition is the separation of an observed time series into estimated components — trend, seasonal, and remainder (with cycle folded into trend at short horizons) — under an assumed structure, most commonly additive (the components sum to the observed value). It is used in this guide as an analytic reading instrument: a way to see which stories a series contains and how large each is, before choosing a forecasting method that must track them. A decomposition is fitted at one declared period; a decomposition at period 7 estimates the weekly pattern and folds everything annual into trend and remainder.

Source: Adapted from Cleveland et al. (1990) and Hyndman and Athanasopoulos (2021).

Table 10.1 collects the components with their StyleCraft instances and — the column that makes the table load-bearing for the rest of the chapter — the instrument that will be responsible for each. The table is a map of Sections 10.5 through 10.9 before the reader gets there: the seasonal-naive baseline handles the weekly pattern by memory, Holt’s method adds a tracked trend, Holt-Winters tracks both, and the three lower rows are the ones no fitted method in the toolkit handles — the annual calendar, which needs an explicit calendar model; the break, which belongs to the analyst; and the noise, which belongs to the interval.

Table 10.1
The components of StyleCraft’s daily revenue series, and what carries each

10.4 Calendar Effects and Structural Breaks

The seasonal component of Section 10.3 deserves a sharper look before any method is built on it, because “seasonality” in marketing data is really a bundle of calendar effects — and most members of the bundle are not carried by a seasonal period at all. Calendar effects are the systematic influences the calendar exerts on a business series: the day-of-week pattern; the fixed-date holidays; the floating retail events — Thanksgiving and the Black Friday weekend it anchors — whose dates move within a window; and paydays, school calendars, and the merchandising rhythms a drops-driven brand imposes on itself. What unites them is knowability: every one of them is already on next year’s calendar. What divides them is whether a fitted seasonal component can carry them, and this is the distinction the draft of this chapter blurred and the audit of it corrected.

The boundary that follows is stated once here and honored everywhere after, because half the chapter’s evaluation design exists to enforce it. This chapter’s Holt-Winters model is fitted at seasonal_periods=7. It tracks the weekly rhythm and nothing else seasonal. StyleCraft’s annual holiday demand is two observations of a 365-position pattern, one of them mid-expansion; it cannot be estimated reliably from that, and this chapter does not pretend otherwise. Annual holiday structure and moving retail events are carried by three instruments instead, all of them the analyst’s: an explicit calendar layer estimated from the prior holiday and aligned on the moving date rather than the calendar date (Code 10.9); documented scenarios that vary the calendar assumption (Code 10.20); and written overrides with owners and review dates (Section 10.9). The seasonal-naive rule does not rescue this either: at a seven-day lag it repeats last week, not last December, and a daily annual seasonal-naive rule would have to reckon with 365- and 366-day calendars and with holidays that move. The chapter does not need dynamic harmonic regression to make this point. It needs to stop implying that a period-7 model carries annual holiday structure — and then to build an evaluation that proves it does not.

The structural break is the calendar effect’s dangerous cousin: an event that changes the series’ underlying behavior rather than decorating it. A seasonal effect visits and leaves; a break moves in. After a break, the level, the trend, or even the seasonal pattern of the series is different — permanently, or at least indefinitely — and the history from before the break describes a process that no longer exists.

definition

Structural Break

A structural break is a point in time at which the data-generating process behind a series changes — a shift in level, trend, or seasonal pattern that persists after the event rather than reverting. History from before a break is evidence about a different regime: models fitted across an unmodeled break blend the two regimes into a description of neither, nor models fitted only on pre-break data will systematically mis-forecast the post-break series. In marketing, breaks are frequently self-inflicted and announced in advance — store openings, price restructurings, channel launches — which makes them knowable to the analyst even though they are invisible to any method that reads only the series’ past.

Source: Adapted from Hyndman and Athanasopoulos (2021).

StyleCraft’s series is a curriculum of breaks by design. Every store opening in the roster added a step to daily revenue: a new level, sustained, with a ramp as the store found its customers — the suburban and flagship step-ups the dataset specification plants explicitly, and the reason Section 10.3’s trend panel shows stairs instead of a slope. The expansion inflection the planning lead must forecast across is exactly this: the cumulative effect of the Wave-4 openings, several recent enough that their ramps are still completing, which means the series’ most recent months are governed by a regime the earlier history only partially describes. Why does history before a break mislead after it? Because every method in this chapter forecasts by assuming continuation — of a level, a trend, a rhythm — and a break is precisely a discontinuation. A model fitted through an unacknowledged opening will read the step as either an alarming trend (and over-extrapolate it) or an outlier (and ignore it); a model fitted only on pre-opening data will confidently forecast the smaller company that no longer exists. Neither model is broken. Both are answering the question they were asked — continue the past — and the past stopped being the operative regime on a date printed, in StyleCraft’s case, in the stores table’s opening_date column.

That last clause is the section’s real payoff, and it earns the chapter’s division of labor its box. Marketing’s structural breaks are unusual among forecasting problems in being disproportionately scheduled: the company decides to open the store, launch the channel, restructure the prices — and writes the date down before it happens. The information exists; it simply does not live in the series. It lives in the stores table, the campaign calendar, the board deck. The analyst who joins that knowledge to the model — through the overrides of Section 10.9, or simply through choosing what history a model is allowed to learn from — is not contaminating the forecast with opinion. She is supplying the one input the method is structurally incapable of finding, and the whole verification design of Code 10.12 is built to make the point unforgettable: the model will be graded, on the known past, on an inflection the reader can see coming and the model cannot.

10.5 Baselines in Time: Naive and Seasonal-Naive

The toolkit proper opens where Chapter 8 taught every predictive project to open: with the opponent. Section 8.7 operationalized baselines as deliberately simple rules whose defeat is the model’s entry fee, and its Table 8.3 introduced the last-value baseline — predict that the previous window repeats — noting it as the incumbent with teeth. This section is that row’s time extension, and one sentence of reminder replaces any re-teach: a baseline is a fully specified prediction rule, graded on the same origins with the same metric as every candidate it opposes. What time adds is a sharper pair of rules and a humbling literature about how hard they are to beat.

The naive forecast is the last-value rule at the grain of the series: tomorrow will equal today. For a multi-day horizon, it extends flat — every future day forecast at the last observed value — which sounds too simple to dignify until one remembers what it encodes: the series’ most recent level, which for a trending, breaking series like StyleCraft’s is often closer to the truth than an average over a history the expansion has outgrown. The seasonal-naive forecast upgrades the memory from yesterday to the rhythm: each future day is forecast by the last observed value from the same position in the seasonal cycle. Its examples must be stated at the right grain, because conflating them is how students end up believing a weekly rule carries an annual pattern. On StyleCraft’s daily series at period 7, the rule is this Saturday equals last Saturday — and that is the only version this chapter uses. On a monthly series at period 12, the rule is this December equals last December. There is no clean daily version of the second: a daily annual seasonal-naive forecast would have to align a 365-day cycle against a 366-day one and would place Black Friday on the wrong day every year, which is exactly the moving-holiday problem of Section 10.4. The rule’s charm, at the weekly period, is that it inherits the entire weekly pattern for free — every stable fixed-period calendar effect carried automatically, at the cost of zero fitted parameters.

definition

Naive and Seasonal-Naive Forecasts

The naive forecast predicts that every future value equals the most recent observed value — the last-value baseline of Section 8.7 applied at the series’ own grain, extended flat across the horizon. The seasonal-naive forecast predicts that every future value equals the most recent observed value from the same position in the seasonal cycle at a declared period: at period 7 on daily data, each Saturday is forecast by last Saturday; at period 12 on monthly data, each December by last December. Both are fully specified, zero-parameter rules, computable from the series alone, and they are the standard opponents of time series forecasting: a fitted model that cannot beat the seasonal-naive rule on an honest evaluation has not yet demonstrated that its fitting bought anything.

Source: Adapted from Hyndman and Athanasopoulos (2021).

Students meet these rules expecting them to be strawmen, and the field’s most instructive empirical tradition exists to correct the expectation. The M-competitions — decades of blind forecasting tournaments across thousands of real business series, most recently the M4 competition’s one hundred thousand — return the same finding with the reliability of a season: simple methods are embarrassingly competitive, many sophisticated entrants fail to beat them, and the average gap between a tuned statistical model and a seasonal-naive rule is far smaller than the gap between either and a bad decision about which history to trust (Makridakis et al., 2020). The lesson is not that fitting is pointless — the best methods do win, and Section 10.6’s family is among them — but that the win is a margin, usually a modest one, and a forecaster who cannot state her margin over seasonal-naive is a forecaster who does not know whether her model is contributing anything. StyleCraft’s series is designed to teach the calibrated version of this humility: the weekly rhythm is strong, so seasonal-naive posts a respectable score, and the labs’ designed expectation is that Holt-Winters beats it by a real but unspectacular margin earned almost entirely where the baseline is blind — the trend the rule cannot climb, since last Saturday is always one week of growth behind this one.

The baseline thread the guide has carried since Chapter 2 is now fully assembled, and it is worth one sentence of signposting: baselines entered as interpretive discipline (Section 2.7), operationalized into opponents (Section 8.7), and here acquire a calendar — and when Chapter 12 gives them visual form as reference lines, the thread will be complete. For this chapter’s purposes the working rule is the one the frame of Code 10.6 declares before any model is fitted: the seasonal-naive forecast, at the weekly period, is the opponent; the naive forecast is reported beside it as the floor; and every exhibit in the deliverable carries the margin over both, in the declared metric, on the declared evaluation — because a forecast’s value to the plan is not its accuracy but its advantage over the free rule the planning lead could have used without hiring anyone.

10.6 Smoothing: Moving Averages and the Exponential Smoothing Family

The baselines remember one value; fitted methods earn their keep by remembering more, and this section builds the chapter’s working family in the order of what each layer adds. The entry instrument is the moving average: replace each day’s value with the mean of a window of days around or behind it. A seven-day trailing moving average of StyleCraft’s revenue erases the weekly sawtooth — each window contains exactly one of each weekday — and what remains is the smoothed level and trend, which is why the moving average is the classical decomposition’s engine and the analyst’s fastest reading tool for “what is the business actually doing under the rhythm.” Two disciplines attach. The window is a dial: seven days removes weekly rhythm, twenty-eight removes most calendar texture, and a window wider than the feature of interest removes the feature — smoothing is deletion, and the analyst should know what she is deleting. And a moving average used for reading is centered or trailing as convenience dictates, but a moving average used inside any forecasting pipeline must be trailing only — a centered window consumes future days, which at forecast time do not exist, and Section 10.11’s audit will treat a centered window in a feature the way Section 8.4 treated any information from beyond the wall.

definition

Moving Average

A moving average replaces each observation with the mean of a fixed window of surrounding observations — trailing (the window ends at the observation) or centered (the window straddles it) — smoothing away fluctuation at periods shorter than the window. A window equal to the seasonal period removes the seasonal pattern entirely, which makes the moving average both the classical instrument for extracting trend and a standing reminder that smoothing is deliberate information removal.

Source: Adapted from Hyndman and Athanasopoulos (2021).

The moving average smooths history; the exponential smoothing family turns smoothing into forecasting, and its governing idea is a graceful improvement on the flat window: weight the past by recency, with weights that decay exponentially, so that yesterday matters most and last quarter still whispers. Simple exponential smoothing maintains exactly one running quantity — the level, an exponentially weighted average of everything seen so far — updated each day by moving a fraction (the smoothing parameter) toward the newest observation. Its forecast is that level, extended flat: simple smoothing is, in effect, a naive forecast with a better memory, appropriate for series with no trend and no season, which StyleCraft’s is not. Holt’s method adds a second running quantity — the trend, the series’ current per-day climb, itself exponentially smoothed — and forecasts by extending the level along the trend: a line, not a flat (Holt, 2004). A damped variant of Holt’s method shrinks that climb as the horizon lengthens, on the empirically well-supported ground that trends estimated from recent data usually flatten rather than continuing forever; it is a candidate in this chapter’s comparison precisely because a company mid-expansion invites over-extrapolation, and whether damping helps is a question for the evidence rather than for taste (Gardner, 2006). Holt-Winters completes the family by adding the third quantity this chapter has been circling: a set of seasonal effects, one per position in the cycle — seven, at StyleCraft’s declared weekly period — each updated when its weekday comes around and applied to the trended level so the forecast carries the rhythm forward (Winters, 1960). Fifty years of practice have kept the family central for the reasons this guide values: it is transparent enough to explain in a meeting, fast enough to refit weekly, and empirically hard to beat on exactly the seasonal, trending business series marketing generates (Gardner, 2006).

definition

Exponential Smoothing, Holt’s Method, and Holt-Winters

Exponential smoothing is a family of forecasting methods that maintain running, recency-weighted estimates of a series’ components, updating each by a smoothing parameter as new observations arrive. Simple exponential smoothing tracks a level only and forecasts it flat. Holt’s method adds a smoothed trend and forecasts along it; a damped variant shrinks the projected trend as the horizon lengthens. The Holt-Winters method adds smoothed seasonal effects at a declared period and forecasts the trended level with the seasonal pattern reapplied — level, trend, and season, each with its own parameter, fitted by minimizing historical forecast error. The declared period is a modeling decision: a model fitted at period 7 carries a weekly pattern and no other.

Source: Adapted from Holt (2004), Winters (1960), and Gardner (2006).

Table 10.2 states the ladder in the form the chapter will use it: what each rung adds, what it costs, and how it fails — because each added component is also an added way to be wrong. The costs compound quietly. Each layer introduces a parameter fitted from history, and fitted parameters are flexibility in exactly Section 8.8’s sense: a trend parameter can mistake a two-week surge for a climb and extrapolate it; seven seasonal effects estimated from a noisy stretch can enshrine noise as rhythm — overfitting’s forecasting costume, caught the same way Chapter 8 caught it, by out-of-sample evaluation. And one failure is shared by the whole family, stated here so Table 10.2 can carry it: every rung forecasts by continuation, so no rung sees a structural break coming, and no rung at period 7 sees a holiday season coming either. The ladder climbs trends and rides weekly rhythms, and walks, at full confidence, straight past the announced opening and into December.

Table 10.2
The smoothing ladder: what each layer adds and what it costs

10.7 Honest Evaluation in Time: Rolling Origins, the Sealed Holdout, and Forecast Accuracy

Section 8.5 closed with a bridge this section now crosses: forecasting cannot randomly split its data, and here is why. The random split of Chapter 8 worked because its wall in time lived inside each customer’s label — features before the snapshot, outcome after — so shuffling customers could not shuffle the future into the past. In a time series, the observations are the time, and a random split shatters the wall into confetti. Hold out a random twenty percent of days and every held-out day sits surrounded by training days: last Tuesday in the training set, next Tuesday in the training set, and the “test” Tuesday between them needing only interpolation — an average of its neighbors — to be predicted with flattering precision. The flattery is structural, not accidental. Adjacent days share trend, season, and even weather; a model graded on interpolation is being graded on a job that never occurs in deployment, because the deployed forecast never knows the future side of the gap. In Section 8.4’s vocabulary, a random split of a time series is leakage by design: every training day after a test day is information from the future, and the resulting score measures how well the future predicts itself — Code 10.11 manufactures the demonstration, and the designed result is a shuffled “error” smaller than any honest forecast achieves.

The repair is the time-aware evaluation: put the wall back where deployment puts it. Train on the past, forecast the future, with every training observation dated strictly before every evaluated observation. The verification is a one-line date-order check — the latest training date must precede the earliest test date — and the check earns permanent membership in this chapter’s audit because it is the forecasting sibling of Section 8.12’s recency range check: a single violated inequality convicts an entire evaluation.

But a single contiguous holdout at the end of the series, which is where most forecasting tutorials stop, is not enough, and the reason is the same reason Chapter 8 needed cross-validation. Comparing four candidate methods on one final block makes that block part of model selection: if Holt-Winters is retained because it won there, the same block cannot also serve as an unbiased final grade. Chapter 8 solved this with k-fold cross-validation inside the training data. Time series solve it with the rolling origin.

definition

Rolling-Origin Evaluation

Rolling-origin evaluation estimates out-of-sample forecast accuracy by repeating one exercise from several successively later forecast origins: at each origin, fit on everything up to that date, forecast a fixed number of periods ahead, record the errors against what happened, and then move the origin forward. Every training set precedes its own test observations, so no evaluation consumes the future, and averaging across origins gives a more stable estimate than one path can. Because each origin produces one error per step ahead, the design also permits accuracy to be reported separately at each forecast horizon. It is the time series counterpart of cross-validation, and it is performed inside the training data so that a final holdout can stay sealed.

Source: Adapted from Tashman (2000) and Hyndman and Athanasopoulos (2021).

The two instruments do different jobs, and this chapter uses both in a fixed order that Table 10.3 states. Rolling origins inside the training data choose the method: naive, seasonal-naive, Holt-Winters, and damped-trend Holt-Winters are compared on the same origins with the same metrics, and the winner is named before anything else happens. Then the method is frozen and one final holdout is opened, once, for the grade. And the final holdout is chosen for a property most tutorials ignore: it must resemble the decision. This matters enough to be the chapter’s largest design commitment. The plan under discussion is a holiday plan. A method graded on eight weeks of late spring has been tested on nothing that matters to it — no November–December demand, no holiday promotions, no Black Friday timing, no annual seasonal effect. So, the sealed holdout in this chapter is November 1 through December 31, 2025, forecast from data ending October 31 — a prior holiday season, forecast from before it, which is the same shape as the actual commission. The result is uncomfortable and instructive: the weekly model, graded on the season it was never told about, misses by roughly twice its rolling-origin error and misses in one direction. That is the point. A holdout that resembles the decision is the only holdout that can expose the gap the decision will fall into.

Table 10.3
The two evaluation layers, and the third window that is not an evaluation at all

With the where settled, the what: forecast accuracy metrics. The dollar metrics transfer whole — MAE and RMSE were built in Section 8.6 and are not re-taught; one sentence of recontextualization suffices. Each evaluated day’s error is actual revenue minus forecast revenue, MAE prices the typical daily miss in dollars, RMSE amplifies the large misses, and the choice between them still follows the decision’s cost structure per Section 2.7. What forecasting adds is the metric Section 8.6 deliberately deferred to this chapter: the percentage form. Planning conversations run on percentages — “we missed the quarter by four percent” — because a percentage travels across scales: it lets the planning lead compare forecast quality across a $9,000 Tuesday and a $14,000 Saturday, across this year and last, across StyleCraft and the benchmarks in her trade association’s survey. The standard instrument is MAPE, the mean absolute percentage error: each day’s absolute error divided by that day’s actual value, averaged, times one hundred.

definition

MAPE (Mean Absolute Percentage Error)

MAPE is the average of a forecast’s absolute errors expressed as percentages of the actual values: for each period, the absolute error divided by the actual, averaged across the evaluated periods and multiplied by 100. It is scale-free, which makes it the planning world’s shared currency for forecast accuracy — and it is undefined at actual values of zero and explosive near them, because a small denominator turns a modest dollar miss into an enormous percentage. MAPE is therefore reported only after its denominators are checked, and it is disqualified for series that touch or approach zero, where the dollar metrics of Section 8.6 must lead.

Source: Adapted from Hyndman and Koehler (2006).

The near-zero failure deserves the paragraph the definition promises, because it is not a curiosity — it is the most common way a competent forecast acquires a headline number that destroys its credibility. Divide by a near-zero actual and the quotient is arithmetic dynamite: Code 10.2 stages the demonstration, where a single storm-closed day of a few hundred dollars converts a forecast whose typical miss is five percent into a reported MAPE in the hundreds — one denominator, not one hundred bad forecasts, produced the number. StyleCraft’s chain-level daily revenue sits comfortably above zero, so MAPE is a legitimate lead metric for this deliverable; but the moment the same pipeline is pointed at a single store’s series, a single channel’s, or a new store’s launch weeks, near-zero actuals arrive and the metric must be benched. This chapter therefore does not leave the check to the analyst’s memory. The MAPE helper of Code 10.1 asserts that every denominator is strictly positive and refuses to return a number otherwise, and the reporting helper prints the minimum denominator and a count of thin days beside every MAPE it displays. Hyndman and Koehler (2006) supply the deeper survey of scale-free alternatives for readers who meet series this guide’s rule disqualifies.

One more quantity completes the evaluation vocabulary, because time gives the frame’s horizon (Section 8.2) a structure it did not have at the customer grain. A forecast is not one claim but a sequence of claims — tomorrow, next week, week eight — and the claims are not equally hard. Uncertainty compounds with each step ahead: trend estimates drift further from the truth, and the world has more time to break. So expected error grows with horizon, and the plan must price it, because the December weeks it cares most about sit at the horizon’s far, foggy end. But the sentence must be said carefully, and the draft of this chapter said it too strongly. Realized error in one evaluation period need not rise monotonically, because each week ahead also differs in promotions, openings, weather, seasonal level, and plain luck — a spike in week seven may be a promotion rather than a consequence of being seven weeks out. This is exactly why horizon accuracy is estimated across several rolling origins rather than read off a single forecast path: with one path, a horizon curve is the history of one two months; with seven origins, each horizon has seven errors, and the pattern begins to mean something. Code 10.10 builds the table that way and prints how many errors sit behind each row.

definition

Forecast Horizon

The forecast horizon is the span of future periods a forecast claims — the distance, in the series’ own grain, from the last observed value to the furthest forecast one. Because uncertainty compounds with each step ahead, forecast uncertainty and expected error generally increase with horizon, although realized error need not rise monotonically in any one evaluation period. Honest evaluation therefore reports accuracy as a function of horizon rather than as a single pooled number, estimates each horizon’s error from several forecast origins rather than one, and states how many errors stand behind each figure. A method’s one-week-ahead grade and its eight-week-ahead grade are different facts, and the decision consumes the one that matches its own lead time.

Source: Adapted from Hyndman and Athanasopoulos (2021) and Tashman (2000).

10.8 Point Forecasts, Prediction Intervals, and Scenarios

Everything so far has graded a forecast as a number meeting a number; this section widens the forecast into what it is, because the widening is the deliverable’s most consequential exhibit. A fitted method, asked about a future Tuesday, does not know a value — under its own assumptions it implies a distribution of plausible values, centered where the components point and spread by what the components cannot carry. The point forecast is a one-number summary of that distribution — its center, essentially — and the summary is legitimate right up until it is mistaken for the claim itself. The claim is the distribution, and the instrument that communicates it is the prediction interval.

definition

Prediction Interval

A prediction interval is a range around a point forecast derived from the forecast distribution the fitted model implies. Under the model and its assumptions, an 80% interval is constructed so that comparable forecast intervals would contain their future observations about 80% of the time. Its width is a claim about uncertainty: wide where the method knows little, growing with horizon, and honest only if the method’s error assumptions are honest. A point forecast reported without an interval conceals exactly the information — the plausible range of the future — that a planning decision most needs.

Source: Adapted from Chatfield (1993) and Hyndman and Athanasopoulos (2021).

The definition’s opening clause carries the qualification that keeps the instrument honest, and this chapter states it twice because it is the sentence practitioners most often drop. An interval is not a statement about the future. It is a statement about the future under the model — its structure, its estimated parameters, its assumption that whatever generated the past will keep generating the future. This chapter builds its intervals by simulation: draw many random future paths forward from the fitted Holt-Winters state, then read percentiles across the paths. That method propagates future disturbances through the fitted states and compounds them across the horizon, which is genuinely the largest part of the uncertainty at short horizons. What it does not do is re-estimate the model’s parameters for each simulated path, so the band is conditional on the parameters that were fitted, and it prices nothing at all about the calendar the model was never told. The working statement, which belongs verbatim in the deliverable: the simulation band reflects future innovation uncertainty under the fitted Holt-Winters model and the propagation of that uncertainty across the horizon; it is conditional on the estimated parameters and may understate total uncertainty when the parameter estimates or the structural assumptions are themselves unstable. The labs make the omission concrete rather than rhetorical: on the sealed holiday holdout the daily band achieves close to its nominal coverage while the season total lands near the top of its range, because the band prices noise faithfully and prices the missing holiday not at all.

Three readings make the interval a working tool rather than a decoration. First, width is honesty, not weakness. An interval that spans plus-or-minus twelve percent of December revenue is not the analyst hedging; it is the method reporting the size of its own ignorance, measured from its own errors — and a competitor who quotes the same future with a bare point number has not reduced the uncertainty, only concealed it. The planning instinct to prefer the confident number is precisely backward, and Section 10.15 will make the ethical version of the argument. Second, intervals fan. Because uncertainty compounds with horizon (Section 10.7), the band widens with each step ahead — the plotted forecast wears a cone, narrow tomorrow and wide in week eight — and the cone’s shape is itself decision-relevant information: it tells the planning lead how much of the plan can be committed early at low risk and how much should wait for the re-forecast. It also tells her when the band has widened past usefulness, which for a period-7 method asked about a date six months out is a real and reportable finding rather than an embarrassment. Third, intervals are themselves gradable — but grading them takes more than one number.

Coverage is the obvious grade: what share of evaluated periods landed inside the band. It is necessary and, alone, badly incomplete, because an arbitrarily wide interval achieves perfect coverage while saying nothing. The draft of this chapter over-read a single coverage figure in both directions, treating a number near eighty as vindication and a number well above it as proof the band was too wide. Neither inference is safe on one short, serially dependent holdout, and the correction is a reporting standard rather than a rule of thumb: report the nominal coverage, the realized coverage, the average interval width, the number of evaluated periods, and coverage by horizon where the periods permit it — then read the set together. Coverage of 94 percent on sixty-one dependent daily observations is a reason to investigate, not a verdict; coverage of 70 percent is likewise a question. Distributional forecasts are ideally evaluated with instruments that reward coverage and sharpness at once, which this guide names and leaves to later coursework (Hyndman & Athanasopoulos, 2021); at this level, the discipline is simply to publish the width beside the coverage so that no one can buy the second by inflating the first.

The interval prices the uncertainty the method can see; the scenario prices the uncertainty it cannot. A scenario is a conditional forecast: the same machinery run under a stated assumption about the calendar the model cannot read — the openings ramp as planned, the openings stall; the holiday behaves like last year’s, it behaves like the weaker of the two years on record. Scenarios and intervals answer different questions and the deliverable needs both: the interval says how wrong the method tends to be when the world stays in regime, and the scenario set says how the forecast moves when the analyst’s declared assumptions do. For this plan, the practice is the three-line table the planning world already speaks — a base case carrying the declared calendar, a downside and an upside carrying the named assumption changes — with the assumptions written beside the numbers, because a scenario whose assumption is unstated is just three guesses wearing a rubric. And the deliverable must say which instrument is the planning range at which horizon. For the eight weeks after the forecast origin, the simulation band is meaningful and the scenarios refine it. For December, six months out, the band has widened past usefulness, and the scenario envelope carries the range — an assumption set, honestly labeled, rather than a graded forecast wearing a graded forecast’s clothes.

10.9 Judgmental Adjustment and the Documented Override

The merchant’s thirty percent has been waiting since Section 10.1, and this section is where the guide keeps its promise about the one legitimate way to use it. The temptation the section disciplines is universal: every statistical forecast, presented, is immediately met by humans who know things — and some of them do. The planning lead knows the fall openings are announced; marketing knows the promotion calendar; everyone in the room knows what date Thanksgiving falls on, which is more than the model knows. A forecasting practice that ignores such knowledge wastes real information; a practice that lets every stakeholder nudge the number produces the laundered consensus Section 10.15 examines. The discipline that threads the needle is the judgmental adjustment, made under rules.

definition

Judgmental Adjustment

A judgmental adjustment is a deliberate, human-made change to a statistical forecast, intended to incorporate information the fitted method could not access — a scheduled event, a structural change, contextual knowledge outside the series. Its legitimacy depends on its provenance and its paperwork: adjustments carrying genuine event knowledge improve forecasts, while frequent small adjustments — optimism, anchoring, target pressure wearing a forecast’s clothes — reliably degrade them, and the empirical record shows the difference is discipline, not intuition.

Source: Adapted from Fildes et al. (2009).

The empirical record deserves its sentence because it cuts both ways with unusual clarity. Fildes and colleagues examined more than sixty thousand forecasts and adjustments inside real supply-chain forecasting operations and found that judgmental adjustments are extremely common; that larger adjustments tended to improve accuracy more than small ones, which mostly added noise; and that positive adjustments were substantially less reliable than negative ones — upward revisions were, in effect, optimism applied with confidence (Fildes et al., 2009). The lesson is not “never touch the model” or “trust the humans” — it is that the touch must clear a bar, and the bar is information the model provably lacks. The openings on the fall calendar clear it: they are structural breaks, announced, with sizes estimable from the earlier openings’ ramps — Codes 10.13 through 10.15 size one. The holiday calendar clears it: the model is fitted at period 7 and cannot carry an annual pattern, so a benchmark estimated from the prior holiday and aligned on Thanksgiving is not opinion but arithmetic the model was structurally unable to do. The merchant’s gifting thesis may clear it: it is a claim about a season the history barely samples, and it can enter as a named scenario assumption with its size argued from occasionwear mix. “The team feels good about Q4” does not clear it, and the discipline exists to say so out loud.

Two implementation refinements separate an override that carries information from an override that carries a mood, and both are lessons the labs learn the hard way. The first is shape. The prose of every planning deck describes new stores as ramping — finding their customers over weeks — and then the arithmetic adds a single flat number from the first day. That is not the stated business process; it assumes the store contributes its steady-state lift immediately, and it will overshoot the first week while undershooting the fourth. An override that matches the process estimates a profile — days 1 to 7, 8 to 14, 15 to 21, 22 to 28 — from prior openings and adds the week-specific figure. The model can stay intentionally simple; what must not stay wrong is the mismatch between what the chapter says the business does and what the code does.

The second refinement is the evidence’s provenance, and it has two halves. The first is contamination control. A before-and-after mean is the natural estimator and an uncontrolled one: the four weeks after an opening differ from the four weeks before it in season, promotions, weather, neighboring openings, and organic trend, and all those lands in the estimate alongside the store. The labs therefore control what can be controlled and disclose what cannot. Equal-length windows keep the weekday mix balanced on both sides. Openings whose windows overlap another opening are excluded, because two steps in one window cannot be separated. Any opening whose evidence window touches November or December is excluded — not merely those whose own date falls there, since a late-October opening measures December in its fourth week just as surely. Promotional days are counted across the pre and post windows together, because a promotion sitting in the before window inflates the baseline and depresses the apparent step exactly as much as one in the after window exaggerates it. Estimates are stratified by store type where the count permits, since a flagship and a resort store have no reason to lift equally. Every surviving opening’s own estimate is printed rather than averaged away, so the reader can see the spread; the override then carries a range rather than a point, which is the only honest form for an estimate built from a handful of observations; and the code asserts that at least one uncontaminated opening survives rather than quietly averaging an empty table.

The second half is which estimate the override is allowed to use, and this is where the chapter departs from the obvious answer. A trend-tracking method fitted on a series full of recent openings has already extrapolated part of the expansion into its own trend term. Adding the raw before-and-after step on top of that adds the same growth twice. So, the labs compute the lift two ways for every surviving opening: the raw step, which is what the series did, and the next step, which is the post-opening actual minus what the fitted method itself forecast for those same days — the part the model demonstrably did not already carry. The net profile is the base adjustment, because it is the only one that cannot double-count. The raw profile is retained as an upper-bound sensitivity and becomes the upside scenario of Section 10.13’s envelope. The gap between the two rows is not noised to be tidied away; it is a measurement of how much of the apparent step the model was already carrying, and the override log records it. Two consequences follow that students should expect rather than be surprised by. A net week-one adjustment can legitimately be negative, which is why the log says the forecast was adjusted rather than raised. And where the net range across openings spans zero, the evidence has not established a lift at all, and the honest instrument is a named scenario rather than a point adjustment.

The instrument of the discipline is the written override, and its form is the point. Every adjustment to the statistical forecast is recorded with six elements: what changed (the periods and the amounts), the direction and size, the evidence (the announced opening; the prior-holiday benchmark; the estimated ramp from earlier openings, with its range and its net-of-trend figure), the owner (who is accountable for this adjustment, by name), the review date (when actuals will grade it), and the baseline preserved (the unadjusted statistical forecast, kept beside the adjusted one so the adjustment’s contribution can be measured after the fact). The log is the forecasting form of disciplines the guide has installed twice already — the verification log of Chapter 4, the analytic specification of Section 2.5 — and it does the same job: it converts an invisible act of judgment into an auditable decision. It also quietly settles a distinction the planning meeting will otherwise blur, and the deliverable states it in one sentence: a forecast is a probability statement about what will happen; a target is an aspiration someone chose; a plan is a commitment of resources — and the moment a target is allowed to edit the forecast that justifies the plan, all three words mean nothing, per the measures-as-targets warning of Section 2.7.

The section closes where the chapter’s decision lives what the forecast is for. Three planning consumers take the same forecast and stress different parts of it. The open-to-buy envelope consumes the seasonal peak and its range — receipts must be committed against a December band, and the asymmetry of Section 2.7 governs which side to lean toward, since unsold inventory becomes markdown budget while stockouts become lost revenue and disappointed stores, and the two costs are not equal. The media pacing plan consumes the weekly shape — when demand arrives, so spend can lead it — and cares more about the seasonal profile’s fidelity than the level’s. And targets consume the forecast last and most dangerously, as the reality check against which aspiration is set; the deliverable’s job there is to keep the words separate. One forecast, three uses, three different tolerances for error — which is why the deliverable of Section 10.13 reports accuracy by horizon and shape, not as one flattering number.

10.10 Forecasting Failure Modes

The toolkit is complete; this section names the three ways it most often fails in marketing practice, each demonstrated on StyleCraft’s own series, and each assigned its repair. They are collected here, just before the AI section, deliberately: every one of them is a failure an assistant will commit fluently, and Table 10.4 is the checklist the audit of Section 10.11 will run.

The first and costliest is extrapolating through a structural break — the failure Sections 10.4 and 10.6 built the vocabulary for. The mechanism: a method fitted on history that ends before (or straddles) a break forecasts the regime it learned, and the regime is gone. The StyleCraft demonstration is Code 10.12: train Holt-Winters with a cutoff just before a Wave-4 opening, forecast across it, and watch a well-fitted, baseline-beating model undershoot the post-opening weeks — not because the fitting failed but because the question was malformed; the model was asked to continue a past the company had already decided to end. The repair is never more fitting; it is the analyst’s calendar (Section 10.4) and the documented override (Section 10.9) — or, where a break sits inside the training window, the harder judgment of which history still describes the current regime, since a model fed pre-break years is being fed evidence about a different company.

The second is overfit seasonality — Section 8.8’s disease at the calendar’s grain — and this chapter’s design turns it from a warning into a demonstrated boundary. The mechanism: seasonal effects are parameters, one per period position, and estimating them from thin repetitions turns noise into rhythm. StyleCraft’s weekly pattern is safe: about a hundred repetitions estimate seven effects generously. Its annual pattern is the trap: two Decembers of history — one of them mid-expansion, one of them containing a store opening — are two observations of a 365-position pattern, and a pipeline coaxed into fitting granular annual seasonality will faithfully reproduce whichever accidents those two Decembers contained, including the promotion spikes and the opening ramps that were never seasonal at all. The repair is structural modesty, and this chapter enacts it rather than recommending it: weekly seasonality is fitted; annual structure is never fitted and is carried instead by an explicit calendar layer estimated from the prior holiday and aligned on the moving date, by documented scenarios that vary that layer across the two years of evidence available, and by overrides with owners. The standing reflex is the one question that catches this everywhere: how many times has the history seen this pattern?

The third is the ignored promotion, the quiet one. The mechanism: promotions move revenue on chosen dates, and a pipeline that never joins the campaigns table misfiles those movements — past spikes read as seasonality (if promotions recur near the same weeks) or as noise (if they do not), and future promotions are absent entirely. The forecast then fails twice: it carries phantom lift into weeks where no promotion is planned, and it misses real lift in the weeks where one is. The StyleCraft demonstration is Code 10.5, where the ten largest decomposition remainders are joined to the campaign calendar and the hit rate is compared against the share of days that are promotional — the null expectation, printed beside the count, so that the exhibit is a comparison rather than an impression. The repair at this guide’s level is the calendar discipline of Section 10.4 and the override machinery of Section 10.9: promotions declared, past ones flagged in the training window’s reading and in the override’s evidence windows, future ones entered as documented adjustments with sizes argued from past lifts.

Table 10.4
The three standing failure modes of marketing forecasting

10.11 AI as a Forecasting Assistant

The division of labor this guide has refined across four chapters — the assistant drafts mechanics, the analyst audits meaning — meets its sternest test in forecasting, for a structural reason worth stating before the failure modes: forecasting is the first task in this guide where the standard tooling’s defaults are wrong. The train_test_split habit, the shuffled cross-validation, the interpolation-friendly evaluation — the general-purpose machine learning reflexes an assistant has absorbed from a million repositories are calibrated for exchangeable rows, and time series data punishes every one of them. An assistant asked to “build a forecast and report accuracy” will, with real probability, deliver a pipeline that shuffles days into a random split (the leakage-by-design of Section 10.7, committed in one default argument), quotes a wonderful error number earned by interpolation, and narrates the result with congratulations. The pipeline will run. The number will be excellent. And the whole artifact will be the vendor button of Section 10.1 rebuilt at home.

The forecasting-specific failure modes, named for the audit. The shuffled split: random or k-fold evaluation applied to ordered data — caught in one line by the date-order check. The reused holdout: even with a genuine time-aware block, a pipeline that compares four methods on that block and then reports the winner’s score on it has spent its own grade; the repair is rolling origins for selection and one sealed block for the verdict, and an assistant will not build that separation unless the prompt demands it. The mismatched season: a holdout chosen for convenience — the last eight weeks, whatever they are — rather than for resemblance to the decision, so a holiday plan is validated on late spring. The invented season: a seasonal period guessed at 12 on daily data, or granular annual seasonality fitted from two repetitions. The silent extrapolation: a fluent forecast straight through the announced openings and straight through December, because no prompt mentioned them and no period-7 method can find them. The benched-metric revival: MAPE quoted on a series with near-zero days, the denominator checks never run. The summed interval: weekly and season-total bounds produced by adding daily bounds together, which is the arithmetic error of Section 10.8 and the one most likely to survive review because the numbers look plausible. The naked point: forecasts delivered without intervals, uncertainty unmentioned. And the method upgrade: asked for a forecast, the assistant reaches for the most impressive machinery it knows — auto-tuned black-box models the student cannot explain — which fails this course’s standing rule that the analyst of record must be able to explain the method to the meeting that funds it, and which usually cannot beat seasonal-naive on two years of daily retail data anyway, per Section 10.5’s M-competition humility.

Where the assistant genuinely helps, use it deliberately, and in forecasting the honest list is long: the groupby-and-reconcile scaffold of Code 10.3; decomposition and plotting code, which is fiddly and perfectly specifiable; the baseline computations and the fixed leaderboard table; the rolling-origin loop itself, which is tedious date arithmetic under frozen rules — exactly what Section 8.9’s discipline makes safe to delegate; error tables by horizon; the path-aggregation arithmetic of Code 10.17, once the analyst has specified which order the operations go in; and the first draft of the interval-communication paragraph, audited hard before it ships. The governing instrument extends the audit lineage one more time: Table 8.5’s five points run unchanged — frame, leakage, split hygiene, baseline, error in decision units — and Table 10.5 adds the four checks forecasting makes necessary. Together they are this chapter’s verification theme made procedural, and the theme’s distinctive move is stated in the box: predict the number before you run the pipeline, because the analyst who has computed the baselines already knows the plausible band, and a result outside the band — in either direction — is a finding about the pipeline, not the future.

Table 10.5
The forecasting supplement to the five-point audit (run with Table 8.5)

ai in practice

The Forecast You Priced Before You Ran It

The two-prompt pattern of Sections 8.11 and 9.11, with this chapter’s twist: the prediction comes first, in writing, before the assistant runs anything.

Step one, the frame as specification: paste the series definition (daily revenue from the certified marketing_daily, reconciled per Code 10.3), the three windows with their dates (rolling-origin selection through October 31, 2025; sealed holdout November 1 to December 31, 2025; planning horizon July to December 2026), the declared metric (MAPE with asserted denominators and minimum printed, MAE beside it), the opponents (seasonal-naive at period 7, naive as floor), the seasonal structure to fit (weekly only; no annual fitting, under any circumstances), the known calendar (openings and promotions in every window, and the date of Thanksgiving in each year touched), and the instruction that closes the escape routes: “Compare candidates on rolling origins inside the selection window only; do not touch the sealed holdout until I name the method; report every candidate and both baselines in one table, by horizon week, with the number of errors behind each figure; produce 80% intervals and aggregate paths before “quantiling” for any weekly or total figure; narrate nothing yet.”

Step two, the prediction: before running, write down the rolling-origin MAPE you expect for all four rows — you can compute the baselines by hand from the series, and the models should land between the seasonal-naive score and the designed noise floor. Then write down, separately, what you expect the sealed holiday holdout to do to that number, and why. Step three, run, then audit: Table 8.5’s five points, then Table 10.5’s four. Step four, only after survival: “Draft the plan paragraph: the December range under each named scenario, the margin over seasonal-naive in dollars, the overrides listed with owners and review dates.” Audit its verbs — a forecast is expected, projected, likely between; it is never will be — and sign the forecast as the analyst of record, interval and all.

Source: Course concept developed for this guide, informed by Hyndman and Athanasopoulos (2021) and Tashman (2000).

10.12 Marketing Interpretation and Managerial Insight

The lab’s outputs are leaderboards, cones, and a log; the commitments run on sentences. This section translates — and, per the guide’s standing practice, it does so partly by exhibiting the wrong managerial readings and correcting them, because forecasting mints a special class of misreading: its numbers sound like promises, its percentages sound like guarantees, and both will be spoken in the planning meeting by people whose only error is treating a probability statement as a commitment.

The first wrong reading is the one the leaderboard invites, and it will be voiced the moment the leaderboard goes up: “The model’s MAPE is 9 percent. So, we’ll be within 9 percent in December — plan accordingly.” Every clause is wrong in a way a section of this chapter has priced, and this chapter’s evaluation design was built so that the correction is a number rather than a caution. A MAPE is an average, not a bound: the evaluated days missed by more and less, and roughly half the misses exceeded the average — the interval, not the MAPE, states the range. It is an average at measured horizons: December sits at the far end of the cone, where Code 10.10’s horizon table showed error running well above the pooled number, and the December weeks deserve the week-seven-and-eight grade. It was measured on the wrong season: the 9 percent came from rolling origins across ordinary spring and summer windows, and the same frozen method, graded on the sealed November–December holdout, missed by close to twice that — in one direction. And a percentage of December is the largest dollar figure of the year: nine percent of the season’s peak weeks is a different commitment problem than nine percent of a quiet April, which is why the deliverable translates every accuracy figure into dollars at the season’s scale before the meeting does it badly. The corrected sentence, worth writing on the leaderboard itself: the method’s rolling-origin error is X percent, its error on the last holiday season it was graded against was Y percent with a systematic undershoot, and the planning range for this December is the scenario envelope, not either percentage.

The second wrong reading is quieter and does more organizational damage: the forecast read as a commitment. “You forecast $1.7 million and we got $1.5 — you were wrong, and planning built the receipts on your number.” The correction has two halves, and the deliverable must make both in advance, not in January. The first half is Section 10.8’s: the forecast was a distribution summarized by a point, the range travelled with it, and a forecast is wrong when actuals fall outside the communicated range more often than the stated rate — not whenever the point misses, because the point always misses. The second half is Section 10.9’s word discipline: the moment the midpoint was copied into the plan as a commitment — and the range quietly deleted in the copying — the organization converted a probability statement into a promise nobody made. The deliverable’s defense is structural: the range travels with the number in every artifact, the scenario table names what would move the outcome to each edge, and the planning translation states explicitly which side of the range each decision leaned toward and why.

The third wrong reading is the one this chapter’s own draft invited, and correcting it is the reason Section 10.1 put a boundary in the brief. “The forecast says $1.7 million, so buy $1.7 millions of product.” A revenue forecast is not a buy. Between the two-sit realized price, category and size mix, gross margin, expected markdowns, inventory already on hand, sell-through, and store allocation — every one of them an assumption, none of them present in a chain-level daily revenue series. What the deliverable can honestly produce is what Code 10.21 produces: an open-to-buy envelope in dollars, built from the forecast by a declared identity whose four terms are visible and vary by scenario, with the assumption the envelope is most sensitive to named. What it cannot produce is a unit plan, and saying so is not a hedge — it is the difference between handing the buying team a number they can interrogate and handing them a number that will be interrogated in January when it is too late. The unit-level forecast is real work at a different grain, with its own baselines and its own holdout, and naming it in the deliverable is the analysis pricing its own upgrade.

The deliverable that survives all three misreadings has a fixed anatomy, assembled entirely from lab outputs. Page one is the frame and the calendar: the series definition with its reconciliation certificate from Code 10.3, the three declared windows with their dates and their different rights per Table 10.3, and the known-calendar list of Section 10.4 — openings, promotions, and the date of Thanksgiving in every year the analysis touches — with each forecast-window item matched to an override or a stated assumption. The rolling-origin leaderboard follows, per Section 8.9’s fixed rules: naive, seasonal-naive, Holt-Winters, and damped Holt-Winters on the same origins, in MAPE with minimum denominator and MAE beside it, with the week-eight column and the bias column shown and the margin over the opponent stated in dollars per day. Then the sealed holdout, once: the frozen method and the frozen method plus the calendar layer, graded on a prior holiday season, with the signed error and the season-total miss — the exhibit that tells the meeting what the model cannot know. Then the horizon table with its error counts. Then the inflection exhibit from Codes 10.12 through 10.15 — both methods undershooting the opening, and the net ramp override closing the shape as well as the level, with the raw profile shown beside it as the upper bound it is — because it is the one exhibit that inoculates the room against the vendor button’s confidence. Then the override log, six elements per entry, with the net-versus-raw distinction in the evidence field and the statistical baseline preserved. Then the interval paragraph: the daily coverage with its width and its period count, the weekly and season-total bands-built path-first, and the sentence about what the band does not price. Then the scenario table, one internally consistent world per row, with both assumption columns, the scenario envelope and the simulation band reported side by side, and a sentence naming which of the two dominates at this horizon. And finally the planning translation: the open-to-buy envelope at each edge with its declared assumptions and its sensitivity, the asymmetry of Section 2.7 argued explicitly — markdown cost against stockout cost, and which side the recommendation therefore leans — the media pacing shape at the declared weekly blocks, the scope statement naming what the envelope is not, and the re-forecast calendar: who reruns this, on what cadence, and which meeting reopens it when October actuals land. A forecast, in this guide’s definition, is all of that; the number alone is just the part that fits in a cell.

10.13 Business Analytics in Practice

This section turns from the fictional case to how forecasting operates in demand planning, disruption management, and media pacing — where its outputs move physical goods and real budgets, and where the hardest-won lessons are about bias, breaks, and the difference between an inaccurate forecast and an unexplained one. The first vignette below draws on the published record of retail forecasting practice; the second draws on the documented industry response to the 2020 disruption; the third is a composite of recurring professional patterns rather than a report about a single named organization.

10.13.1 Demand Planning, S&OP, and the Bullwhip

The first vignette is demand planning and sales-and-operations planning, the institutional home of forecasting in most product companies, and the place where a forecast most directly becomes money. The cadence is recognizable from this chapter’s own decision context, run at industrial scale: a statistical baseline forecast is generated by item and location; planners and commercial teams adjust it in a consensus process; and the adjusted forecast drives purchase orders, production schedules, and inventory positions months ahead of the demand it predicts. Retail adds two complications the published research treats as central, and this chapter has been building toward. Demand is hierarchical — the same forecast is wanted at chain, category, store, and SKU grain, and the accuracy that is achievable falls sharply as the grain gets finer, which is exactly why this chapter’s deliverable declares a chain-level dollar envelope and refuses the unit plan. And promotions dominate the error: at store-SKU grain, a large share of the variance a retailer most needs to anticipate is generated by its own promotional calendar rather than by any underlying seasonal process (Fildes et al., 2022b).

Two failure patterns dominate the genre’s scar tissue. The first is bias: consensus processes tilt optimistic, because targets, incentives, and enthusiasm all push one direction — which is why mature demand-planning organizations track signed error as seriously as absolute error, and audit whether human touches to the statistical forecast add accuracy at all, the practice the judgmental-adjustment evidence of Section 10.9 directly supports (Fildes et al., 2009). The second failure compounds down the supply chain: the bullwhip effect, in which modest variability or bias in demand signals amplifies as each tier — retailer, distributor, manufacturer — orders against its own forecast of the tier below, so a small forecast error at the shelf becomes a large swing at the factory (Lee et al., 1997). The practice lesson lands directly on this chapter’s machinery: a forecast is an input to other people’s decisions, its errors propagate with leverage, and the disciplines that look pedantic at the desk — bias tracking, preserved statistical baselines, documented adjustments — are, at supply-chain scale, the difference between a bad quarter and a warehouse of markdowns.

10.13.2 The Season the Models Never Saw

The second vignette is the structural break at civilization scale. In the spring of 2020, pandemic lockdowns delivered a break no training window contained, and forecasting systems across retail experienced it as a controlled experiment nobody wanted: models trained on years of stable seasonality met a world where grocery demand spiked, apparel and travel collapsed, and e-commerce absorbed years of channel shift in weeks. The instructive part is not that the forecasts were wrong — everything in Section 10.4 says they had to be — but what the field concluded afterward, and the postscript to the retail forecasting literature is unusually direct about it. Established forecasting approaches, tuned on stable history, degraded badly; the period exposed how thin the research base was on forecasting under abrupt structural change and on instability in general; and the practical responses that helped were the ones that shortened the reach of history and widened the statement of uncertainty (Fildes et al., 2022a).

Read those responses against this chapter’s repairs, because they are the same list applied under fire. Teams shortened training windows and down-weighted the broken period, which is the which-history-still-applies judgment of Section 10.10. They re-forecast more frequently, shrinking the horizon because expected error grows with it. They widened intervals and communicated ranges where they had shipped points, because the honest statement of uncertainty had become impossible to avoid. And they shifted weight from fitted history to structured judgment — documented adjustments carrying information no model possessed, which is Section 10.9 operating as the primary engine rather than the garnish. The durable lesson survived the emergency: the pandemic period now sits in every retailer’s training history as a permanent monument to the principle that a model cannot announce its own obsolescence. The humans who knew the world had changed had to say so, in writing, and the organizations that could do that quickly were the ones whose forecasting practice already separated the statistical baseline from the judgmental layer.

10.13.3 Media Pacing as Weekly Micro-Forecasting

The third vignette is small, fast, and the one most students will live first; it is a composite of a practice pattern rather than a named case. Every performance marketing team runs a version of it: a monthly or quarterly budget, spend and revenue accruing daily, and a standing question — where will we land? — answered continuously from partial-period actuals. The naive arithmetic everyone starts with (“we are 40 percent through the month at 35 percent of budget, so we will underspend”) is a forecast, and usually a bad one, because it ignores exactly the structure this chapter taught: day-of-week rhythm, so a month’s spend is not linear in days; intra-month and moving-holiday seasonality, since auction prices and conversion shift toward month-end and toward the retail events whose dates move; and self-inflicted breaks — a creative refresh, a bid-strategy change, a competitor entering the auction. Teams that pace well run, in miniature, the full discipline: a seasonal-naive expectation of the period’s shape, a short-horizon projection updated daily, explicit adjustments for the promotions and launches on the calendar, and a range rather than a point when the finance partner asks for the landing. The practice lesson is proportion: the horizon is days and the stakes are one budget line, but the habits — baseline first, calendar declared, adjustment documented, range communicated — are the identical habits the holiday plan needed, which makes pacing the training ground where a junior analyst can practice this chapter weekly with fast feedback.

10.13.4 In Your First Analyst Job

In your first analyst job, these vignettes compress into one expectation, and it is this chapter’s closing thread: businesses forgive inaccurate forecasts and punish unexplained ones. Every organization that plans has been missed by its forecasts — the veterans in the room have watched seasons collapse and surprises rally, and they know the future is not fully knowable. What they do not forgive is the miss that cannot be explained: the number with no method behind it, the point with no range around it, the adjustment nobody wrote down, the December surprise that turns out to have been an announced opening the forecast silently ignored, or a holiday the model was never told about. The analyst who ships the full deliverable — method, opponent, grade on a season that resembles the decision, ranges, calendar, override log, review date — will still be wrong, routinely, by roughly the amount the holdout promised. But she will be wrong in a way the organization can learn from, and her forecasts will be trusted more each season, because the trust was never really about accuracy. It was about accountability, and accountability is a document, not a talent.

10.14 Ethics, Forecast Accountability, and the Laundered Guess

The Business Analytics in Practice section ended on accountability as a professional survival skill; this section examines it as an ethical obligation, extending the guide’s running discussions — data use (Section 1.13), problem framing (Section 2.12), measurement design (Section 3.13), cleaning as editorial power (Section 4.14), honest summarization (Section 5.14), differential treatment of segments (Section 6.16), causal language (Section 7.15), acting on predictions about people (Section 8.15), and fairness across groups (Section 9.15) — to what forecasting adds: a number about the future that other people will stake resources, jobs, and judgment on, produced by a process only the analyst can see.

The chapter’s ethics angle has a blunt name: the laundered guess. Laundering, in the financial sense, is passing something questionable through a process that makes it look clean, and a guess is laundered the same way: a number of uncertain provenances — a gut feeling, an unvalidated model, a target wearing analytical clothes — is passed through the apparatus of precision (decimals, a chart, a planning template) and emerges looking like knowledge. The vendor button of Section 10.1 launders structurally: a method nobody can inspect emits a number to the dollar, and the precision of the digits impersonates the reliability of the process. The consensus meeting launders socially: a forecast nudged upward by target pressure leaves the room wearing the statistical model’s name. And the analyst launders personally in three specific ways this chapter has now named. She launders when she ships a point forecast without its range, because she possesses the holdout errors and has chosen to present a confidence the method does not have. She launders when she quotes a grade earned on a convenient window as though it applied to the decision’s own season. And she launders when she reports a quantity her analytical frame cannot produce — a unit buy derived from a revenue forecast by assumptions nobody wrote down. In every case the mechanism is the same: uncertainty that someone knew about was stripped somewhere between the method and the meeting, and the audience downstream inherited a certainty nobody held.

Why is this an ethical matter rather than a technical one? Because of whom bears the cost when the point forecast is treated as a promise. The planning lead who commits receipts against the laundered number owns the markdowns; the store teams whose staffing and bonus targets were set from its own the shortfall; the finance partner who guided the quarter on it owns the surprise. The analyst’s exposure — an awkward January meeting — is routinely the smallest in the chain, which is precisely the asymmetry that makes the disclosure obligatory: the person most able to state the uncertainty is the person least exposed to the cost of concealing it. The interval is therefore not a statistical garnish but the ethical instrument of the trade — the sentence “between X and Y, eight times in ten, under the model, and here is what the model does not know” is the honest transfer of what the analyst knows to the people who will bear what she does not — and the guide’s position is correspondingly plain: a point forecast delivered to a resource decision without its range, by an analyst who could have computed one, is a misrepresentation, however accurate the point later proves. Being right is not a defense; the audience was entitled to the uncertainty and got confidence instead.

Accountability, the angle’s other face, is what makes the obligation livable rather than paralyzing, and its instruments are the chapter’s own. The forecast is signed — the analyst of record, per Chapter 1, extends through the forecast’s assumptions and overrides, each owned by name with a review date. The uncertainty is communicated at the decision’s grain — the interval with its width and period count, the horizon table, the scenario assumptions, and the statement of which instrument is the planning range at which horizon — so that leaning optimistic or conservative becomes the decision-maker’s explicit, priced choice per Section 2.7 rather than the analyst’s hidden one. The overrides are documented so that judgment is auditable, per Section 10.9, and reviewed against actuals so the organization learns which judgments to keep. The scope is declared, so that a dollar envelope is not silently consumed as a unit plan. And the forecast is kept separate, in writing, from targets and plans, because the measures-as-targets dynamic of Section 2.7 has a forecasting form with a long industrial history: the moment the forecast is graded on whether it matches the aspiration, it stops informing the aspiration, and the company is navigating by a compass wired to its own hull.

10.15 Chapter Summary

This chapter carried Part II’s predictive discipline into time, and the move cost more than a change of index. The data structure came first: a time series is ordered, its order carries the information, and the exchangeability that let Chapters 8 and 9 shuffle customers freely died at the door — StyleCraft’s daily revenue had to be manufactured from the fact table by a deliberate grain change, reconciled against its source on a complete calendar with the zero-fill rule stated rather than inherited from a library default, and read before it could be modeled. Reading meant decomposition: the series resolved into trend, seasonality, cycle, and noise, with the store-opening dates drawn onto the trend panel and the ten largest remainders joined to the campaign calendar against a printed chance expectation, so that “promotions explain the spikes” arrived as a comparison rather than an impression.

The chapter’s organizing correction is a boundary. Calendar effects are knowable in advance, but only stable fixed-period patterns are carried automatically by seasonal memory; moving holidays, promotions, and openings must be supplied explicitly. A Holt-Winters model at seasonal_periods=7 tracks a weekly rhythm and nothing else seasonal, so StyleCraft’s annual holiday structure — two observations of a 365-position pattern, one of them containing a store opening — is not fitted at all. It is carried instead by an explicit calendar layer estimated from the prior holiday and indexed on days from Thanksgiving rather than on calendar dates, by scenarios that vary that layer across the two years of evidence that exist, and by written overrides with owners and review dates. The toolkit was then built as a ladder with the opponent on the bottom rung: the naive and seasonal-naive rules extended Chapter 8’s baselines into time and proved, per the M-competitions’ standing humility, brutally hard to beat; moving averages smoothed the rhythm away for reading; and the exponential smoothing family — level, then Holt’s trend, then a damped variant, then Holt-Winters’ weekly seasonal effects — added tracked components one at a time, each purchasing power with flexibility.

Honest evaluation was rebuilt for ordered data in two layers rather than one. Random splits were convicted of leakage by design — the shuffled evaluation grades interpolation, a job no deployed forecast performs — and replaced by rolling origins inside the training data, which chose the method across seven origins and fifty-six-day horizons, and by one sealed holdout opened once for the frozen method. The holdout was chosen for a property most tutorials ignore: it had to resemble the decision, so it is a prior holiday season forecast from before it, and the result is the chapter’s most useful number. The method that won on ordinary windows missed the holiday by roughly twice its selection score, systematically low, which is not a defect but the measured size of the calendar the model was never given. MAPE joined the dollar metrics as planning’s shared currency and acquired a guard in code — denominators asserted, minimum printed beside every figure — after the miniature showed one storm-closed day converting a five-percent forecast into a headline of two hundred sixty. Error was reported by horizon across several origins rather than one path, with the number of errors behind each row printed, and the claim was stated at the strength the evidence supports: expected error and uncertainty grow with horizon, while realized error in any one sample need not rise at every step.

The forecast was then widened from a number into a distribution, and the widening carried two corrections. Weekly and season-total intervals are built by aggregating each simulated path first and taking quantiles second, because the sum of seven daily tenth percentiles is not the tenth percentile of the seven-day total — and the lab measured how wrong the shortcut is on real paths and on paths whose dependence had been shuffled away, to show that the size of the error cannot be known without the paths. And the band was qualified rather than trusted: it is conditional on the fitted parameters, it prices future innovation and horizon compounding, and it prices the missing calendar not at all — which the sealed holdout demonstrated by achieving near-nominal daily coverage while the season total landed near its upper bound. Coverage was reported with width, period count, and horizon detail, so that no one can buy a coverage figure by inflating a band.

The human hand was disciplined rather than banished. Judgmental adjustment earned its legitimacy exactly where it carries information the history cannot — the announced opening, the moving holiday, the drafted promotion — and paid for it with implementation as well as paperwork: a ramp profile rather than a flat step, because the prose said the stores ramp; equal weekday windows, excluded overlapping openings, excluded holiday-season openings, a promotion-day cap, stratification by store type, every surviving estimate printed, and a range carried forward instead of a decimal; plus a net-of-trend diagnostic that asks how much of the apparent step the model was already extrapolating, whose instability is itself the finding. Then the six-element written override, whose deepest payoff arrives in January when the preserved baseline lets the organization compute what each judgment contributed.

The verification theme ran through all of it: the labs forecast the known past across a programmatically selected opening, watched the well-fitted model and the free rule fail identically in direction, and closed both the level and the shape with a documented ramp. And the deliverable finally reached the decision without overclaiming: a Q4 scenario envelope with its assumption columns, a horizon caution stating that the graded horizon was fifty-six days and December sits three times further out, an open-to-buy envelope built from a declared identity whose every term varies by scenario, a media pacing shape at declared seven-day blocks, and a scope statement naming what a chain-level revenue forecast cannot produce — the unit buy, which is separate work at a separate grain. Twenty numbered cells carried the argument into executable form, and the failure modes, the AI supplement, the interpretation, the practice mirror, and the ethics section circled one insistence from different distances: a forecast is a probability statement with a method, an opponent, a grade on a season that resembles the decision, a range with a stated scope of what it prices, and a signature. Everything less is a guess, laundered.

One forward note redeems a pointer the smoothing section deferred: the moving averages built here return in Chapter 14 as visual analytics, computed live on the Tableau canvas, where their window arithmetic — learned here — is the difference between reading them and merely displaying them.

Looking ahead, the chapter’s machinery answers what will happen: December’s range, the season’s shape, the opening’s step. It is silent, by construction, on the question the planning meeting will ask next — what caused it, and what would happen if StyleCraft acted differently? The forecast can say the suburban stores’ revenue will step up; it cannot say the expansion caused the step, or what December would have looked like without it, because everything in this chapter learned from a single unrepeated history in which StyleCraft did only one thing. The next chapter takes up the machinery that can answer such questions: experiments, A/B testing, and causal evidence — random assignment as the instrument that manufactures the comparison history never provides, the known true lift the dataset ships so the analysis can grade itself, and the evidence hierarchy that finally places everything Part II has built — descriptions, models, forecasts — on the ladder of what each one licenses a marketer to claim.

10.16 Exercises for Practice

The following exercises practice the chapter’s main habits: build and reconcile the series before modeling it, read the decomposition before choosing a method, name the opponent before fitting, separate method selection from the grade, choose a holdout that resembles the decision, report MAPE with its denominators and error by horizon across origins, aggregate intervals in the right order, document every override with its six elements, and declare the limits of the planning unit. They are organized into two groups. Core chapter practice is the required path and should be completed by every student.

10.16.1 Core Chapter PracticeExercise 10.1 Concept Check

Answer each in two or three sentences, in your own words.

• State what makes a time series a different data structure from the customer feature table and name the discipline from Chapter 8 that survives the change and the one that does not.
• Distinguish seasonality from a cycle and both from a structural break, with one StyleCraft instance of each.
• Explain why history from before a structural break misleads a model after it, and why this is not a defect the model can be tuned out of.
• A colleague says that because Christmas and Black Friday are known in advance, a Holt-Winters model will carry them. Correct the claim precisely, distinguishing knowable from carried, and name the three instruments that carry the effects the model does not.
• Write the naive and seasonal-naive forecasting rules for a daily series with weekly rhythm and explain why this guide states the December-repeats-December version only for monthly data.
• State what Holt’s method adds to simple exponential smoothing, what the damped variant adds to Holt, and what Holt-Winters adds to both — and name the new failure each addition makes possible.
• Explain why a random train/test split flatters a forecasting model, using the word interpolation, and state the one-line check that catches the flattery.
• Explain why comparing four candidate methods on the final holdout spends the grade and describe the two-layer design that fixes it.
• This chapter’s sealed holdout is a prior holiday season rather than the last eight weeks of the file. Give the reason, and state what a late-spring holdout would and would not have established about a holiday plan.
• A holdout MAPE of 6 percent was computed on a series whose minimum value is $180 against typical values near $11,000. May the MAPE lead the deliverable? Explain, citing the standing rule, and describe what the chapter’s mape helper would have done.
• Explain why the sum of seven daily 80% intervals is not the weekly 80% interval and state the correct procedure in one sentence.
• A stakeholder reads an 80% prediction interval as “the analysts are only 80% sure of their work.” Correct the reading in two sentences, including what it would mean for the interval to be wrong and what the interval does not price at all.

Exercise 10.2 Baselines and MAPE by Hand

Using only the fourteen-day miniature of Code 10.1 (no code): rebuild both forecast rules day by day and verify every figure— the naive rule’s errors and its MAE of 1,300.00 and MAPE of 12.79 percent; the seasonal-naive rule’s errors and its MAE of 514.29 and MAPE of 5.25 percent. Identify the two week-two days on which the naive rule’s error is largest and state, in calendar language, what each miss has in common. Show that all seven seasonal-naive errors share a sign, compute their mean, and write one sentence on which row of Table 10.1 the bias reveals. Then rebuild the storm scenario: verify the MAE of 1,471.43 and the MAPE of 261.52 percent, show which single term produced the explosion, and state what value the storm Tuesday would have had to take for the chapter’s guarded mape helper to refuse to compute at all. Close with the two-sentence honest summary of the storm week that the standing denominator rule requires.

Exercise 10.3 Read the Calendar

For each series below, list the calendar effects and structural breaks a forecaster should declare before fitting, classify each as living in the training window, the forecast window, or both, and state — per Section 10.4’s obligation — how each forecast-window item should be discharged: by a fitted seasonal period, by an explicit calendar layer, by a documented override, by a stated scenario assumption, or by scope exclusion. Where you answer, “fitted seasonal period,” you must also state how many repetitions of the pattern the history contains.

• StyleCraft’s daily chain revenue, forecast from July 2026 through the end of December 2026, given the stores table, the drafted holiday promotion calendar, and a board deck listing two further openings under consideration for October.
• A single Wave-4 resort store’s daily revenue, opened five months ago, forecast eight weeks ahead.
• The email channel’s daily attributed revenue, where sends occur three days a week and revenue on non-send days is frequently near zero.
• Weekly chain revenue for the four weeks around Thanksgiving, forecast one year ahead, where the analyst proposes to use last year’s same-week values.

Exercise 10.4 Spot the Failure

Each scenario below contains at least one failure from this chapter, or none. Name it, cite the section or table row, and state the repair.

• An analyst evaluates a daily revenue model with five-fold cross-validation, reports a MAPE of 2.1 percent, and notes the model “performs far better than the seasonal-naive rule’s 7 percent.”
• A pipeline’s feature list includes a centered 14-day moving average of revenue.
• An analyst compares naive, seasonal-naive, and Holt-Winters on the final eight weeks of the series, keeps Holt-Winters because it wins, and reports its error on those eight weeks as the model’s out-of-sample grade.
• A forecast for October–December is fitted on the full two-year history, and the analyst notes with satisfaction that the model has “learned the holiday season” from granular day-of-year seasonal effects.
• A team forecasts across an announced flagship opening using Holt-Winters, and when the forecast undershoots, concludes the smoothing parameters need retuning.
• A weekly forecast table reports 80% bounds obtained by summing the daily 80% bounds within each week. The analyst notes that the difference “is only a few percent anyway.”
• An analyst reports the eight-week forecast as a weekly table of point values with a footnote that “actual results may vary,” and separately keeps a spreadsheet of the intervals “to avoid confusing the audience.”
• A planning lead fits the forecast, finds it 9 percent below the CEO’s announced growth target, and raises the seasonal effects until the gap closes, noting the December numbers “looked low anyway.”
• A deck reports a chain revenue forecast of $1.7 million for Q4 and, on the next slide, a recommended purchase quantity of 34,000 units. No other assumption appears anywhere in the deck.

Exercise 10.5 The Full Forecasting Deliverable

Complete Labs 10.1 and 10.2 on the certified files and assemble the planning deliverable per Section 10.13: the frame-and-calendar page with the reconciliation certificate and the three declared windows with their different rights; the rolling-origin leaderboard for all four candidates with MAPE, minimum denominators, MAE, the week-eight column, the bias column, and the margin over the opponent in dollars per day; the sealed-holdout grade, opened once, for the frozen method and the frozen method plus the calendar layer, with signed error and season-total miss; the horizon table with its error counts and the December-distance sentence; the inflection exhibit with both methods’ bias across the programmatically selected opening and the ramp override’s measured contribution to level and to shape; the override log with all six elements per entry, including the range and the net-of-trend figure in the evidence field; the interval paragraph with coverage reported beside width and period count, the weekly and season-total bands built path-first, and the sentence stating what the band does not price; the scenario table with its assumption columns and the horizon caution; and the planning translation with the open-to-buy envelope, the asymmetry of Section 2.7 argued explicitly, the media pacing shape, the scope statement, and the re-forecast calendar. Submissions are graded on the calendar page, the separation of selection from grade, the override log, the interval communication, and the scope statement as heavily as on the models.

Exercise 10.6 AI Forecast Audit

Give an AI assistant the raw marketing_daily file and this deliberately loose prompt: “Forecast our daily revenue for the next eight weeks — maximize accuracy.” Before reading the response, write your predicted leaderboard band per the AI in Practice box: the seasonal-naive MAPE you expect (compute it yourself) and the band an honest model should land in. Then audit the response in writing with Table 8.5’s five points followed by Table 10.5’s four: locate the split and run the date-order check; determine whether method selection and the reported grade happened on the same data; determine whether any feature or smoother consumed post-cutoff data; reconstruct what the pipeline assumed about openings, promotions, and the holiday; check every metric’s denominators and every forecast for its missing interval, including whether any weekly or total interval was built by summing daily bounds; and compare the headline claim against your predicted band, classifying any excess as pipeline, not prophecy, until traced. Re-prompt with the full frame and compare leaderboards. Conclude with two sentences on which supplement point caught the most serious problem.

10.17 Glossary of Terms

This glossary includes only the terms introduced in this chapter. Each definition is tied to the sources used in the chapter rather than added for decoration.

Calendar effect. A systematic influence the calendar exerts on a business series — day-of-week rhythm, fixed-date holidays, moving retail events, paydays, self-imposed merchandising rhythms. All are knowable in advance; only stable fixed-period patterns are carried automatically by seasonal memory, and the rest must be supplied explicitly (adapted from Hyndman & Athanasopoulos, 2021).

Cycle. A recurring rise and fall in a time series without a fixed, calendar-locked period — economic or fashion waves whose timing cannot be read from a calendar, distinguishing them from seasonality (adapted from Hyndman & Athanasopoulos, 2021).

Decomposition. The separation of an observed time series into estimated trend, seasonal, and remainder components under an assumed structure, most commonly additive; used in this guide as a reading instrument that shows which stories a series contains and how large each is, at one declared period (adapted from Cleveland et al., 1990; Hyndman & Athanasopoulos, 2021).

Exponential smoothing. A family of forecasting methods that maintain running, recency-weighted estimates of a series’ components — level alone in simple smoothing, level and trend in Holt’s method, level, trend, and season in Holt-Winters — updating each by a fitted smoothing parameter as observations arrive (adapted from Holt, 2004; Gardner, 2006).

Forecast horizon. The span of future periods a forecast claims, in the series’ own grain. Because uncertainty compounds with each step ahead, forecast uncertainty and expected error generally increase with horizon, although realized error need not rise monotonically in any one evaluation period; honest evaluation reports accuracy by horizon, estimated across several forecast origins, with the number of errors behind each figure stated (adapted from Hyndman & Athanasopoulos, 2021; Tashman, 2000).

Holiday calendar layer. This guide’s name for an explicit, analyst-supplied adjustment that carries annual and moving-holiday structure a fitted seasonal period cannot: a ratio of actual to statistical forecast measured across a prior holiday season and indexed by days from the moving anchor date rather than by calendar date. Estimated from one or two observations, it is an assumption with a stated range and a review date, not a fitted component (course concept developed for this guide, informed by Hyndman & Athanasopoulos, 2021).

Holt-Winters method. The seasonal member of the exponential smoothing family: smoothed level, smoothed trend, and smoothed seasonal effects at a declared period, forecast forward as a trended level with the seasonal pattern reapplied. A model fitted at period 7 carries a weekly pattern and no other (adapted from Winters, 1960; Gardner, 2006).

Judgmental adjustment. A deliberate human change to a statistical forecast intended to incorporate information the fitted method could not access; legitimate when carrying genuine event knowledge and documented as a written override, and reliably accuracy-destroying when small, frequent, and optimistic — with upward adjustments substantially less reliable than downward ones (adapted from Fildes et al., 2009).

MAPE (mean absolute percentage error). The average of absolute forecast errors expressed as percentages of the actual values; scale-free and planning’s shared currency, but undefined at zero actuals and explosive near them, so it is computed only after its denominators are asserted, reported with its minimum denominator, and disqualified for series that approach zero (adapted from Hyndman & Koehler, 2006).

Moving average. The replacement of each observation with the mean of a fixed window of surrounding observations, trailing or centered; a window equal to the seasonal period removes the seasonal pattern, making the moving average both a trend-extraction instrument and a deliberate act of information removal. Inside a forecasting pipeline it must be trailing only (adapted from Hyndman & Athanasopoulos, 2021).

Naive forecast. The rule that every future value equals the most recent observed value — the last-value baseline of Section 8.7 at the series’ own grain, extended flat across the horizon (adapted from Hyndman & Athanasopoulos, 2021).

Open-to-buy envelope. This guide’s name for the planning unit a chain-level revenue forecast can honestly produce planned receipts at retail, computed as planned sales plus planned ending inventory plus planned markdowns minus inventory on hand, converted to dollars of buying authority by a declared cost complement. Every term is either forecast or declared and varies by scenario, and the envelope is a dollar figure rather than a unit plan (course concept developed for this guide).

Prediction interval. A range around a point forecast derived from the forecast distribution the fitted model implies under the model and its assumptions, an 80% interval is constructed so that comparable intervals would contain their future observations about 80% of the time. Its width grows with horizon, it is gradable by realized coverage reported beside width and period count, and it is conditional on the fitted parameters — silent about structure the model was never given (adapted from Chatfield, 1993; Hyndman & Athanasopoulos, 2021).

Rolling-origin evaluation. Estimation of out-of-sample forecast accuracy by repeating a fit-and-forecast exercise from several successively later forecast origins, so that every training set precedes its own test observations and accuracy can be reported separately by horizon; the time series counterpart of cross-validation, performed inside the training data so a final holdout can stay sealed (adapted from Tashman, 2000; Hyndman & Athanasopoulos, 2021).

Sealed holdout. A contiguous block of the known past, withheld from every selection decision and consulted once for the frozen method. In forecasting it carries an additional obligation the customer-grain version did not: it should resemble the decision, in season and in horizon, because a method graded on an unlike window has not been tested on the risk the decision runs (course concept developed for this guide, informed by Tashman, 2000).

Seasonal-naive forecast. The rule that every future value equals the most recent observed value from the same position in the seasonal cycle at a declared period — this Saturday forecast by last Saturday on daily data at period 7, this December by last December on monthly data at period 12 — reproducing one cycle of the pattern forward at zero fitted parameters; this guide’s declared opponent for forecasting models (adapted from Hyndman & Athanasopoulos, 2021).

Seasonality. A repeating pattern in a time series with a fixed and known period, recurring at a constant number of observations apart — days of the week, months of the year — and therefore carryable by seasonal memory. Moving holidays are calendar effects but not seasonality in this sense (adapted from Hyndman & Athanasopoulos, 2021).

Structural break. A point in time at which the process generating a series changes — a persistent shift in level, trend, or seasonal pattern — after which pre-break history describes a regime that no longer exists; in marketing, frequently self-inflicted and announced in advance (adapted from Hyndman & Athanasopoulos, 2021).

Time series. A sequence of observations of the same quantity at successive, usually regular, points in time, in which the temporal ordering is part of the data’s meaning and must be preserved because observations may exhibit serial dependence, trend, seasonality, or structural change (adapted from Hyndman & Athanasopoulos, 2021).

Trend. The long-run direction of a time series — the smooth underlying path once short-run fluctuation is set aside; in StyleCraft’s series, organic growth compounded by the expansion’s stair-step contributions (adapted from Hyndman & Athanasopoulos, 2021).

10.18 References

Chatfield, C. (1993). Calculating interval forecasts. Journal of Business & Economic Statistics, 11(2), 121–135. https://doi.org/10.1080/07350015.1993.10509938

Cleveland, R. B., Cleveland, W. S., McRae, J. E., & Terpenning, I. (1990). STL: A seasonal-trend decomposition procedure based on loess. Journal of Official Statistics, 6(1), 3–73.

Fildes, R., Goodwin, P., Lawrence, M., & Nikolopoulos, K. (2009). Effective forecasting and judgmental adjustments: An empirical evaluation and strategies for improvement in supply-chain planning. International Journal of Forecasting, 25(1), 3–23. https://doi.org/10.1016/j.ijforecast.2008.11.010

Fildes, R., Kolassa, S., & Ma, S. (2022a). Post-script—Retail forecasting: Research and practice. International Journal of Forecasting, 38(4), 1319–1324. https://doi.org/10.1016/j.ijforecast.2021.09.012

Fildes, R., Ma, S., & Kolassa, S. (2022b). Retail forecasting: Research and practice. International Journal of Forecasting, 38(4), 1283–1318. https://doi.org/10.1016/j.ijforecast.2019.06.004

Gardner, E. S., Jr. (2006). Exponential smoothing: The state of the art—Part II. International Journal of Forecasting, 22(4), 637–666. https://doi.org/10.1016/j.ijforecast.2006.03.005

Holt, C. C. (2004). Forecasting seasonals and trends by exponentially weighted moving averages. International Journal of Forecasting, 20(1), 5–10. https://doi.org/10.1016/j.ijforecast.2003.09.015

Hyndman, R. J., & Athanasopoulos, G. (2021). Forecasting: Principles and practice (3rd ed.). OTexts. https://otexts.com/fpp3/

Hyndman, R. J., & Koehler, A. B. (2006). Another look at measures of forecast accuracy. International Journal of Forecasting, 22(4), 679–688. https://doi.org/10.1016/j.ijforecast.2006.03.001

Lee, H. L., Padmanabhan, V., & Whang, S. (1997). Information distortion in a supply chain: The bullwhip effect. Management Science, 43(4), 546–558. https://doi.org/10.1287/mnsc.43.4.546

Makridakis, S., Spiliotis, E., & Assimakopoulos, V. (2020). The M4 Competition: 100,000 time series and 61 forecasting methods. International Journal of Forecasting, 36(1), 54–74. https://doi.org/10.1016/j.ijforecast.2019.04.014

Seabold, S., & Perktold, J. (2010). statsmodels: Econometric and statistical modeling with Python. In Proceedings of the 9th Python in Science Conference (pp. 92–96). https://doi.org/10.25080/Majora-92bf1922-011

statsmodels developers. (2026a). statsmodels.tsa.seasonal.seasonal_decompose [Software documentation]. https://www.statsmodels.org/stable/generated/statsmodels.tsa.seasonal.seasonal_decompose.html

statsmodels developers. (2026b). statsmodels.tsa.holtwinters.ExponentialSmoothing [Software documentation]. https://www.statsmodels.org/stable/generated/statsmodels.tsa.holtwinters.ExponentialSmoothing.html

statsmodels developers. (2026c). statsmodels.tsa.holtwinters.HoltWintersResults.simulate [Software documentation]. https://www.statsmodels.org/stable/generated/statsmodels.tsa.holtwinters.HoltWintersResults.simulate.html

statsmodels developers. (2026d). Development release notes for the forthcoming 0.15.0 release [Software documentation]. https://www.statsmodels.org/dev/release/version0.15.0.html

Tashman, L. J. (2000). Out-of-sample tests of forecasting accuracy: An analysis and review. International Journal of Forecasting, 16(4), 437–450. https://doi.org/10.1016/S0169-2070(00)00065-0

Winters, P. R. (1960). Forecasting sales by exponentially weighted moving averages. Management Science, 6(3), 324–342. https://doi.org/10.1287/mnsc.6.3.324