P7 · Time Series & Forecasting
decomposition, classical + ML forecasting, anomaly detection
Order is information: read a series before forecasting it, beat the naive baselines honestly, backtest without leakage and spot the anomalies — the temporal instincts P5's shuffled world never needed.
Every method in the Traditional-ML pillar quietly assumed one thing: that the rows of your table are exchangeable — shuffle them and nothing is lost. A sales history, a server's error counts, a language learner's daily review scores: shuffle any of these and you have destroyed exactly the thing worth modelling. Time series is what machine learning becomes when order is the signal — when yesterday genuinely predicts today, when December is not July, and when evaluating a model by mixing the future into its training data produces numbers that are confident, excellent and wrong.
The first track teaches you to read a series before daring to forecast it: splitting a line into trend, seasonality and noise; measuring how strongly a series remembers itself (autocorrelation); and recognising when a series is stable enough for the classical machinery to apply at all. Everything is worked on small seeded series whose behaviour you can check by hand.
The second track forecasts — starting from the baselines every honest model must beat. Naive repetition and seasonal repetition sound trivial and are embarrassingly hard to outperform; a forecasting effort that loses to them is negative work. From there: exponential smoothing, a gentle reading of ARIMA's three dials, and the reframe that turns forecasting into a supervised-learning problem P5 already taught you. The track closes with the discipline that holds it all together — rolling-origin backtesting, the leakage traps that fake skill, and residual-based anomaly detection.