Recap and final exam

Ten modules to go from a raw daily series to a deployed 28-day forecast with intervals. Here is the course condensed, then the threads that run through it.

The course at a glance

Module	The essential point
1. What makes a series different	Rows are ordered; a shuffled split fakes any metric. Split chronologically, always
2. Decomposition	Trend + seasonality + noise; STL for one period, MSTL for weekly and yearly
3. Stationarity	ADF test then difference; ACF and PACF pick the ARIMA orders
4. Baselines	Naïve, seasonal-naïve, moving average, drift — the first row of the table, MASE's denominator
5. ARIMA and SARIMA	Diagnostic-driven fit; SARIMA beats seasonal-naïve by 14 %; residual whiteness before shipping
6. Smoothing and Prophet	Holt-Winters as short-series workhorse; Prophet for business calendars and interpretability
7. Features and regressors	Lagged targets, calendar cycles, holidays; known-future versus unknown-future is the pitfall
8. Deep approaches	LSTM and Transformers pay off on many related series with rich features; not on a single short one
9. Rolling validation and metrics	One test window lies; per-horizon MAE and MASE, honest interval scoring
10. Project: forecast with intervals	Quantile LightGBM at 6.4 MAE, three-sentence uncertainty story, drift-triggered retraining

The threads running through the course

A shuffled split fakes every metric. Module 1's 3.1 MAE from train_test_split(shuffle=True) was a mirage; the honest number on the same model was 11.6. Every claim in this course rests on a chronological split with a gap for the horizon. When a colleague reports a forecasting metric, the very first question is not "which model?" but "how did you split?".

Baselines are the whole ranking. The seasonal-naïve at 9.8 units per day is the number every other model in the results table is measured against. That baseline is also the denominator of MASE; without it, no metric here would be comparable across stores or across models. Skipping baselines is how teams spend a quarter chasing a Transformer that turns out to be worse than a two-line pandas snippet.

The horizon shapes the model. SARIMA is trained to minimize one-step-ahead error and compounds at horizon 28. LightGBM with direct multi-step or per-horizon models can specialize. Deep models with a wide output head produce all 28 outputs at once. A model excellent at "predict tomorrow" is not automatically excellent at "predict the month".

Known-future versus unknown-future regressors. Planned promotions, holidays, day-of-week: known ahead, safe to use. Weather, competitor moves, current-day flags: unknown ahead, either forecast them first or leave them out. The seemingly small "we'll figure that out at inference" comment is the most common cause of production failure in forecasting projects.

Uncertainty is the deliverable. A point forecast forces the business to guess its own uncertainty. Quantile regression at $\tau = 0.1, 0.5, 0.9$ turns three simple LightGBM fits into an interval whose historical coverage the team can quote. That interval, not the point, is what turns a forecast into a procurement decision.

Decision tree: which method for which situation

• Very short history (under 200 points): exponential smoothing or SARIMA; deep models will overfit.
• Single series, clean weekly cycle, no external effects: seasonal-naïve, SARIMA if you need intervals; the marginal improvement of anything deeper is not worth the cost.
• Business calendar (holidays, campaigns) matters more than autoregression: Prophet, communicated to the analyst team through plot_components.
• Many features, one to few series: gradient boosting on a supervised table, quantile regression for intervals.
• Many related series, rich features, long history: temporal Transformer (TFT if you have known-future regressors), or a foundation model as a zero-shot floor.
• Every setup: rolling backtest for the metric, per-horizon report, an interval alongside the point.

The final exam

The exam has 40 questions covering the ten modules: reading a naïve split as a leak, decomposition and STL, ADF and differencing, ACF/PACF-based order picking, SARIMA diagnostics, Prophet's real strengths and its silent failures, calendar features and the known-future rule, when deep learning earns its cost, MAPE's trap at zero and MASE as the right default, quantile intervals and their coverage.

Several questions are situations to diagnose: a model whose test-set MAE looks too good to be true, a MAPE returning inf, a Prophet forecast whose 80 % interval covers only 68 % of the truth, a LightGBM that beats an LSTM by 5 % on one origin and loses on the next. It is judgment that is assessed, not the recitation of formulas.

On success, your certificate of completion is issued immediately; its number is verifiable by any third party on the platform.

Before you start

Take the table above and, for each row, ask yourself "how would I see that I am wrong here?". If you can say why a shuffled split invalidates every metric, why MAPE breaks at zero, why a good coverage matters more than a small interval, and when deep learning is not worth its cost, you are ready. Good luck!

Final exam

Ready to validate this course?

40 questions drawn at random from the course bank · passing score 70% · verifiable PDF certificate issued immediately on success.

Start the exam

You need to be signed in to your InSkillML account with an active subscription. You can also start the exam from My courses.