Recommended reading

M3 Competition (Makridakis et al., 2000) — 3,003 time series, monthly and quarterly. Theta won. This is where simple methods proved they could beat ARIMA.

M4 Competition (Makridakis et al., 2020) — 100,000 series. An ES-RNN hybrid won, but top non-ML methods were Theta and simple exponential smoothing variants.

M5 Competition (Makridakis et al., 2022) — Walmart hierarchical demand. LightGBM-based ensembles dominated. The winning solution was an equal-weighted blend of 6 LightGBM variants.

M6 Competition (2022–2023) — Financial forecasting. Competitor accuracy on price/return forecasting clustered near luck. Forecast accuracy did not correlate with investment outcomes.

Forecasting: Principles and Practice (Hyndman & Athanasopoulos, 3rd ed.) — The best free textbook on time-series forecasting. Available at otexts.com/fpp3. Covers ETS, ARIMA, and modern methods.

Time Series Analysis (Hamilton, 1994) — The graduate-level reference for ARIMA, GARCH, and state-space models. Dense but authoritative.

statsforecast (Nixtla) — Fast, production-grade implementations of Naive, SES, Holt, Theta, ARIMA, and more. The recommended starting point for classical methods.

prophet (Meta) — Easy deployment of trend + seasonality models with changepoints and holidays. Good for fast iteration on messy operational data.

lightgbm + mlforecast (Nixtla) — Gradient boosting for panel forecasting. The mlforecast wrapper handles feature engineering (lags, rolling means, date features) cleanly.