Time Series Forecasting
Predicting what comes next in a sequence over time — where simple methods beat sophisticated ones for forty years, and only recently stopped.
When not to use it
- When the drivers aren't in the series. If the future depends on a competitor's decision, the history doesn't contain it.
- On a short series, with deep learning. 36 monthly points is not a dataset. Use ETS or ARIMA.
- Without a naive baseline. You don't know if your model works until you know what doing nothing scores.
- When you need a decision and report a point. Without intervals, you've hidden the only part that mattered.
Reach for something else instead
- Seasonal naive — the baseline. Sometimes it's also the answer.
- ETS / ARIMA — decades old, instant, competitive on single short series.
- Gradient boosting with lag and calendar features — the practical default for business forecasting.
- Scenario planning — when the process isn't stable, forecasting is the wrong frame entirely.
Sources & further reading
- Makridakis, Spiliotis & Assimakopoulos (2018), The M4 Competition: Results, findings, conclusion and way forward — the record of ML underperforming statistics, stated by the people who ran it.
- Makridakis et al. (2022), The M5 competition: Background, organization, and implementation — where gradient boosting finally won.
- Hyndman & Athanasopoulos, Forecasting: Principles and Practice — the free textbook; still the best practical reference.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Random train/test splits. The model learns from the future, the score is fiction.
- Skipping the naive baseline because it feels beneath the project.
- Reaching for deep learning on one short series. It needs many related series to have anything to learn from.
- Ignoring known future events. A simple model that knows about holidays beats a complex one that doesn't.
- Reporting point forecasts without intervals, which discards the uncertainty that made it a decision.
At a glance
FieldApplied AI
Baseline to beatseasonal naive
Neverrandom splits
The inflectionM5, 2020, gradient boosting
Why ML laggedmost business series are short
Most undervalued outputprediction intervals
DifficultyBeginner
Flashcards for this concept
Question
—
Answer
—
1 / 4
Often compared with
Statistical vs. ML forecasting — statistics wins on one short series; ML wins across thousands of related ones. The dividing line is how much data you actually have.