Forecasting
A track of P7 · Time Series & Forecasting.
From naive baselines to ARIMA and lag-feature regression — forecasting methods fitted honestly, backtested without leakage, and used to spot anomalies.
Forecasting has a humbling open secret: "tomorrow will look like today" is brutally hard to beat. This track starts there — naive and seasonal-naive baselines, and the MASE metric that scores every later model against them — because a forecasting effort that loses to repetition is negative work, and you should be able to prove it.
From that floor the track climbs: exponential smoothing (yesterday matters more than last month, encoded as decaying weights), a gentle reading of ARIMA's three dials, and the reframe that turns forecasting into the supervised learning you already know from P5 — lag features in, regression out, with one crucial new rule about what honest evaluation means when data has an arrow of time. The final module is that rule made rigorous: rolling-origin backtesting, the catalogue of leakage traps that manufacture fake skill, and residual-based anomaly detection — the analytical cousin of the production monitoring P13 owns.