Part Overview
Long before neural networks, statisticians built a remarkably complete theory of forecasting, and it still wins on many real problems. Part II teaches that classical toolkit not as history but as live, competitive practice. These six chapters cover the methods a working forecaster reaches for first: stationarity diagnostics, the autoregressive family, seasonal and exponential-smoothing models, multivariate and volatility models for finance, state-space filtering, and the detection of anomalies and regime changes. Every one of them has a modern descendant in later parts, so learning them well pays compound interest.
The chapters build outward from the single series to the system. Chapter 3 establishes the statistical language: random processes, stationarity, autocorrelation, and decomposition. Chapter 4 changes coordinates to the frequency domain, where periodicity, filtering, and even modern frequency-mixing forecasters become transparent. Chapter 5 is the heart of univariate forecasting, from moving averages through ARIMA and ETS to Prophet, with the model-selection discipline that keeps them honest. Chapter 6 scales up to many interacting series with VAR, cointegration, and the GARCH volatility models that finance depends on. Chapter 7 introduces the state-space view and the Kalman filter, then frames filtering as recurrent inference, the exact idea the temporal thread will pick up again as the RNN and the structured state-space model. Chapter 8 closes the part with anomaly and change-point detection, the safety net of any deployed temporal system.
The finance running dataset leads here: returns and volatility motivate ARIMA, GARCH, and cointegration, while the industrial sensor stream anchors the anomaly and change-point material. Two arcs of the temporal thread begin in this part and resurface much later: the Kalman filter of Chapter 7 returns as the RNN (Part III) and the structured state-space model (Chapter 13), and the spectral analysis of Chapter 4 returns inside frequency-domain deep forecasters (Chapter 14). Treat Part II as the set of baselines every neural model in the book must beat.
Random processes, stationarity, autocorrelation, and decomposition: the shared vocabulary of every classical method.
Viewing a series as a sum of waves, where periodicity, filtering, and modern frequency mixers all become clear.
From moving averages to ARIMA, ETS, and Prophet, with the model selection and diagnostics that keep them honest.
VAR, cointegration, dynamic factors, and the GARCH family for many interacting series and their volatility.
Hidden states, the Kalman filter, particle filters, and HMMs, framed as the recurrent inference the temporal thread returns to.
Spotting the unexpected and the moment a regime shifts, the safety net of any monitored temporal system.
Where This Part Leads
The classical models of Part II set the baselines and the intuitions; the book now makes them learnable. Part III: Temporal Deep Learning builds directly on this foundation: the autoregressive structure of Chapter 5 reappears as linear attention, the spectral analysis of Chapter 4 returns inside frequency-domain forecasters, and the Kalman filter of Chapter 7 is rediscovered first as the RNN and then as the structured state-space model.