AutoML for Forecasting
Forecasting looks like ordinary prediction until you try to automate it. The target moves in time, yesterday's values are your best features, and shuffling rows — the first thing most machine-learning tooling does — destroys exactly what you are trying to learn. AutoML for forecasting is automated model building reshaped around that constraint.
What changes under the hood
- Validation becomes rolling — the system trains on a window and tests on the period straight after it, then slides forward. This backtest is the only honest score.
- Features come from the calendar and the past — lags, rolling means, day-of-week, holidays, and the gaps a sales series carries when a store was shut.
- The shortlist widens — ARIMA, exponential smoothing and Theta compete against gradient-boosted trees on lag features and deep sequence models, because no family wins on every series.
Where you meet it
Demand planning across thousands of products is the canonical case: nobody hand-tunes a model per item, so tools such as AutoGluon-TimeSeries, Nixtla's StatsForecast, Darts, or the forecasting mode inside Vertex AI and Azure AutoML fit candidates per series, rank them on backtest error, and ensemble the survivors.
Why it matters
The win here is coverage rather than cleverness. A plain seasonal-naive baseline beats a careless model on a short, noisy series, and an automated system that backtests properly will tell you so instead of shipping a flattering number — which is the failure mode that makes hand-built forecasts expensive to trust at scale.
AutoML for forecasting applies automated model search to time-ordered data, where rows cannot be shuffled. The system generates lag, rolling and calendar features, scores candidates with rolling backtests instead of random folds, and ranks statistical models such as ARIMA against tree ensembles and deep sequence models. It matters because forecasting problems arrive as thousands of parallel series — one per product, store or sensor — far more than anyone can tune by hand.
Predicting next month's sales is not the same job as sorting photos of cats. Time matters: what happened last week is the clue to what happens next, so you cannot jumble the past into a random pile.
AutoML for forecasting is software that builds those time-based predictions for you. It invents the useful clues itself — last week's figure, the same week last year, whether a holiday falls in the window — tries lots of methods, and checks each one the fair way: hide the recent past, predict it, see who came closest. Shops and factories use it because they need a forecast for every item on the shelf, not just one.