STOP 31 / 31A new problem. A sensible next step.
Layout
STOP 31 · In practice

A new problem. A sensible next step.

EXPLORE~12 min
Back to route map
  • You have a sales dataset and someone asks for a forecast.
  • Where do you begin?
  • Not with ARIMA orders or a neural network.
  • Start with the decision, understand the data, build a baseline, and let honest tests tell you what to try next.
  • This is a reusable roadmap—not a ladder you must climb to the most complicated model.
A forecasting workflow: concepts and Northstar examples. Illustrative Northstar examples, not measured business results.
ConceptWhat it meansNorthstar example
First define what a useful forecast means.
  • Agree on the target, frequency, level of detail, horizon, and cost of being wrong.
  • For Northstar: daily sales over the next six weeks to support inventory decisions, including the Black Friday period.
  • Check stockouts: observed sales may be below true demand.
  • Northstar first specifies the decision: daily staffing, a 28-day purchase order, or monthly finance planning.
  • Each implies different useful horizons and tolerances.
Explore before fitting; establish a baseline before tuning.
  • Plot the training history.
  • Audit missing dates, outliers, zeros, returns, and structural changes.
  • Inspect weekly and annual patterns; use STL or MSTL as exploratory tools, not standalone forecasting models.
  • Then benchmark naïve, seasonal naïve, and a carefully event-aligned last-year forecast where history permits.

Inspect Northstar's missing days, stockouts, weekly rhythm, and Black Friday timing before choosing a simple event-aware baseline.

Design an honest test, then diagnose the misses.
  • Choose rolling origins and the real business horizon before comparing candidates.
  • Keep a final chronological test period untouched during tuning when enough history permits.
  • Inspect validation error by horizon, weekday, and holiday, alongside bias and interval coverage.
  • Residual plots and ACF suggest what the model missed; they do not replace future-data evaluation.

Test future periods from several historical origins, then ask whether Northstar's misses are bias, omitted events, missing seasons, or residual dependence.

Add the component that the evidence calls for.
  • A changing level suggests SES; a trend suggests Holt or damped ETS; one seasonal rhythm suggests seasonal ETS.
  • Multiple seasonalities suggest comparing MSTL + ETS with Fourier regression, possibly with ARIMA errors.
  • Prophet provides another trend/seasonality/holiday pipeline to test.
  • Remaining serial dependence suggests ARIMA/SARIMA after appropriate stationarity checks.
  • Moving events and promotions need explicit known-ahead features.
  • These are alternatives or combinations—not mandatory successive stages.
  • Add a calendar feature for a missed event or an error model for remaining memory.
  • Do not climb a mandatory model-complexity ladder.
Keep complexity only when it earns its place.
  • Compare candidates on the same origins, horizons, and metrics.
  • Consider an ensemble or machine learning when data and the use case justify them, not because a simple method is unfashionable.
  • Prefer the simplest reliable option with useful intervals and manageable operating costs.
  • Refit, publish assumptions, monitor performance, and keep the baseline as a fallback.
  • If a simple baseline best supports Northstar's decision after honest tests, keep it.
  • Refit, monitor, and revisit the choice when the business changes.
define → audit → baseline → backtest → diagnose → targeted candidate → compare ↺ → deploy & monitor
FIG 31.1

Your roadmap for a new forecasting problem

WHAT THIS SHOWS

Open one stage at a time and map a new Northstar forecasting problem to a decision, honest test, and targeted candidate.

The question it answers: What evidence should determine the next step—not a mandatory complexity ladder?

Select a step to read its explanation right underneath. Only one step stays open at a time.

  1. WHAT TO DO

    Write a one-paragraph forecasting brief before touching a model: what quantity, for whom, at what granularity, how far ahead, and what action depends on it? Decide whether you need individual days, the horizon total, or both.

    AT NORTHSTAR

    Northstar: daily net sales by category for the next 42 days, refreshed weekly. The inventory team also needs plausible high-demand scenarios before Black Friday. Clarify how sales, prices, and availability relate to the units they must stock.

    LEAVE THIS STEP WITH

    An agreed target, information cut-off, horizon, metric, and business owner.

↺

No reliable improvement? Keep the simpler method, or return to audit and diagnosis. You do not have to try every model family.

THE BRANCH AT STEP 6

What should I try next?

Pick the pattern you observed. This suggests an experiment—not a guaranteed winner.

KEEP AS THE BENCHMARK
Weekly naïve plus a separate event-aligned last-year benchmark
CANDIDATE TO TEST
Compare calendar/Fourier regression, Prophet, and MSTL + ETS with event handling

Weekday and annual Fourier features plus event-relative windows form one candidate; DHR adds ARIMA errors if useful. Prophet provides a trend/seasonality/holiday alternative. MSTL + ETS separates multiple seasons but needs deliberate event adjustment and add-back. Compare their complete pipelines under the same validation protocol.

Watch out: Neither smooth Fourier waves nor a fixed seasonal slot reliably aligns a sharp moving holiday by itself. Tune within time-aware validation, and use scenarios for unknown future drivers.

A PRACTICAL FIRST PASS FOR NORTHSTAR

Agree on the 42-day decision → audit sales and availability → benchmark weekly and event-aligned naïve forecasts → evaluate rolling 42-day forecasts → inspect weekday and Black Friday misses → test calendar/Fourier regression with explicit event windows → add ARIMA errors only if useful → compare, refit, and monitor.

STL/MSTL helps explore components. ETS is a forecasting family. Fourier terms represent smooth seasonality. They solve different problems and are not mandatory stages in one fixed sequence.

Fine print & assumptions

Decision guide: validate each candidate on the actual business decision and future-data tests.

CHECK YOURSELF · CHOOSE AN ANSWER

A weekly ETS model does well on ordinary days but repeatedly underpredicts Black Friday. What is the most useful next experiment?

Your learning note

Private · just for you