Time Series Forecasting with Machine Learning

A premium, complete and hands-on course on modern time series forecasting, updated for 2026. You will start from what makes temporal data different, decompose series into trend, seasonality and noise, master autocorrelation with the ACF and PACF, test for stationarity, and run a disciplined exploratory analysis and preprocessing pipeline. You will learn to evaluate forecasts honestly with MAE, RMSE, MAPE, sMAPE, WAPE and MASE, and to validate with temporal cross-validation instead of leaking the future. From strong classical baselines and seasonal-naive benchmarks you move to ARIMA, SARIMA and exponential smoothing (ETS, Holt-Winters), plus the Croston, SBA and TSB methods for intermittent demand, then reframe forecasting as supervised learning with lag, rolling and calendar features to train gradient-boosting models with XGBoost and LightGBM, build global cross-learning models across thousands of series, and tune everything with disciplined, leakage-free hyperparameter searches. You will build deep learning forecasters (LSTM, TCN, N-BEATS, N-HiTS), understand Transformers for time series (Informer, PatchTST, TFT) and the 2026 wave of foundation models (TimesFM, Chronos, Moirai) for zero-shot forecasting, produce calibrated prediction intervals with quantile and conformal methods, handle multivariate and hierarchical problems with reconciliation, and master the production toolkit (statsmodels, Prophet, sktime, statsforecast, Darts) plus deployment, monitoring and retraining. Every concept is paired with real, correct Python code, and the course keeps a strong focus on uncertainty, data privacy and the responsible use of forecasts. Includes a comprehensive final assessment.