Skip to main content

Time series forecasting

From statistical methods to deep models, benchmarked against solid baselines and validated with a rolling origin — on a real 28-day demand forecast.

Course Duration: 6h​

What You'll Learn​

  • Diagnose why a shuffled split silently invalidates every metric on a time series
  • Decompose a daily series into trend, seasonality and residuals with STL, and read the residuals
  • Test stationarity with ADF, choose orders of differencing, read ACF and PACF plots
  • Build the naïve, seasonal-naïve, moving-average and drift baselines that any complex model must beat
  • Fit an ARIMA or SARIMA model with statsmodels, diagnose its residuals, produce forecasts with intervals
  • Use Holt-Winters and Prophet for their real strengths — and know where they break
  • Turn a series into a supervised table with lags, calendar features and holiday flags, then train gradient boosting
  • Train an LSTM and a temporal Transformer on the same task and know when depth actually pays off
  • Set up a rolling-origin evaluation with the right metrics: MAE, RMSE, MASE and the traps of MAPE at zero
  • Deliver a demand forecast with quantile intervals and a plan for retraining

Prerequisites​

  • Mathematics for AI (course 03): means, variances, autocorrelation
  • Supervised learning (course 04): baselines, cross-validation, overfitting
  • Recurrent networks (course 11) is useful but not required — module 8 recaps the pieces we need

Course Modules​

  1. What makes a time series different from other data
  2. Decomposition: trend, seasonality, noise
  3. Stationarity, differencing, autocorrelation
  4. Baseline models you should never skip
  5. ARIMA and SARIMA
  6. Exponential smoothing and Prophet
  7. Calendar features and external regressors
  8. Deep approaches: LSTM and temporal Transformers
  9. Rolling validation and suitable metrics
  10. Project: demand forecasting with intervals

The thread running through the course​

A single business use case runs through every module: daily demand at a pharmacy chain. Four years of sales per day and per store, with weekly and yearly seasonality, national holidays, occasional promotions, and a horizon fixed at 28 days for procurement. Each module adds one row to a shared results table so that the reader sees, from module to module, which method actually helps and which one only trades in complexity. The final module compares them side by side and picks the model that gets deployed — with quantile intervals, because a point forecast that ignores its own uncertainty is the shortest path to over-ordering or a stockout.

Assessment and certificate​

The course ends with a 40-question exam covering every module. It is judgment that is assessed — reading a naïve split that leaks the future, choosing SARIMA orders from an ACF plot, knowing why MAPE misbehaves at zero, deciding when a deep model is worth its cost. On success, a certificate of completion is issued immediately; its number is verifiable on the platform.

Free courses, by contrast, end with a 5-question quiz and a preview of the certificate, without certification.