Skip to content
Time Series & Forecasting

ML Forecasting & Evaluation

When ML Beats ARIMA

9 min+85 XPStep 1/11

What ML actually buys you

Classical models work one series at a time, with linear machinery and a handful of parameters. ML buys three things they can't touch:

  • Nonlinearity and interactions — 'hot and weekend and holiday' combinations a linear model can't express
  • Exogenous features at scale — price, promotions, weather, events, hundreds of columns deep
  • Global models — one model trained across all your series at once, so a new SKU with two weeks of history borrows patterns from thousands of siblings

In practice, gradient-boosted trees (LightGBM, XGBoost) dominate tabular forecasting — the M5 retail competition, 42,000 Walmart series, was won with boosting plus ruthless feature engineering. Deep models (DeepAR, N-BEATS, Temporal Fusion Transformers) shine when you need probabilistic forecasts at scale.

insight

The M5 result was a wake-up call: the winning approach wasn't a cleverer ARIMA — it was a global LightGBM model feasting on good features. Architecture matters less than what you feed it.