Notes
🔍

Time Series AutoML

Forecasting looks like ordinary prediction until you notice that the data has a direction. Yesterday can inform tomorrow, but never the reverse — and most of the machinery of general-purpose AutoML quietly assumes otherwise. Time series AutoML is automated model building rebuilt around that one constraint.

What changes
Four things work differently from the tabular case:

  • validation becomes backtesting — rolling or expanding windows that always train on the past and test on the future, never shuffled folds;
  • features are generated from history: lags, rolling means, differences, calendar and holiday flags, and covariates known in advance;
  • the search spans statistical and learned models together — ARIMA, exponential smoothing and Theta alongside gradient-boosted trees and deep forecasters;
  • the forecast horizon and the series frequency become part of the problem definition rather than an afterthought.

Tools you would reach for
AutoGluon-TimeSeries fits a menu of statistical and neural forecasters and ensembles them by backtest score; sktime and StatsForecast automate model and order selection within the classical family; managed cloud forecasting services wrap the same idea. Auto-ARIMA — automatic order selection by information criterion — is the oldest automated forecaster still in serious use, and a baseline worth beating before anything fancier is justified.

The failure it prevents
Random cross-validation on a time series leaks the future into training and yields a model that scores beautifully and forecasts badly. Getting that right is most of what separates time-series AutoML from pointing a tabular tool at a date column.

Time series AutoML automates forecasting model building for data ordered in time: it generates lag, rolling and calendar features, searches statistical and machine-learning forecasters together, and validates by backtesting on rolling windows instead of shuffled cross-validation. That temporal discipline is the point — random validation leaks future information and yields forecasts that score well and fail in production. The forecast horizon is fixed as part of the problem.

Predicting next month's sales is not like guessing someone's age from a photograph. With sales, the order matters: you can learn from January to predict February, but never the other way round.

Time series AutoML is automated model-building that respects that rule. It builds clues out of the past — last week's figures, the same month a year ago, bank holidays — then tries many forecasting methods and checks each one by pretending to stand at an earlier date and predicting what actually came next. That honest test is the whole reason this flavour exists: shuffle the dates and the computer accidentally peeks at the answer.