Applied Time Series, Part 3: Advanced Topics — State Space, Regimes, and Modern Forecasting Practice

The final course of the three-part Applied Time Series sequence, taking graduates of Part 1, "Applied Time Series: Statistics and Econometrics of Data with Memory" (applied-time-series), and Part 2, "Applied Time Series II: Multivariate Models, Cointegration, and Dynamic Econometrics" (applied-time-series-2), to the working frontier of practitioner forecasting and dynamic modeling. Both prior courses are hard prerequisites: the Box-Jenkins workflow, exponential smoothing, honest rolling-origin evaluation against naive benchmarks (Part 1), and HAC inference, VARs, error correction, GARCH, structural breaks, and the Diebold-Mariano test (Part 2) are used freely from the first lab. Part 3 has three arcs. The first is structure: the state-space framework and the Kalman filter unify everything fitted so far and unlock missing data, signal extraction, time-varying parameters, Bayesian structural time series, nowcasting, and — the practitioner payoff — counterfactual causal impact analysis (intervention models, CausalImpact, synthetic control) and regime-switching models (SETAR, Markov switching) that turn Part 2's one-off structural breaks into recurring, forecastable states. The second arc is modern forecasting practice on hard real-world data: multiple seasonality at hourly frequency (dynamic harmonic regression, MSTL, TBATS, and an honest audit of Prophet), intermittent and count demand (Croston-family methods, INGARCH, and why Part 1's metrics break on zero-heavy series), machine learning with global gradient-boosted models and the leakage discipline the M4/M5 competitions taught, and deep learning plus zero-shot foundation models (DeepAR-style networks, N-BEATS/N-HiTS, Chronos, TimesFM, Moirai) evaluated with the same benchmark skepticism Part 1 installed on day one. The third arc is delivery: probabilistic forecasting done properly — quantile losses, CRPS, calibration, and conformal prediction intervals with finite-sample guarantees — and hierarchical forecasting with MinT reconciliation, closing with a capstone that runs a full production-style pipeline on a retail hierarchy against benchmarks from all three courses. Labs run in both Python (statsmodels, sktime, mlforecast, neuralforecast, hierarchicalforecast, arch) and R (fable/fpp3, KFAS, bsts, CausalImpact, tsintermittent), deliberately using whichever ecosystem is strongest for each method and translating between them. The organizing theme: a forecast is a probability statement, a model is a hypothesis about structure, and the practitioner's job is to choose, combine, and audit both at scale.

  • Level: advanced
  • Duration: 10 weeks
  • 11 sections
  • Price: $50

Topics: state space models, Kalman filter, Kalman smoother, local level model, unobserved components, structural time series, signal extraction, missing data, time-varying parameters, Bayesian structural time series, spike-and-slab, nowcasting, dynamic factor model, MIDAS, intervention analysis, interrupted time series, CausalImpact, synthetic control, counterfactual forecast, Markov switching, regime switching, SETAR, threshold autoregression, smoothed probabilities, multiple seasonality, dynamic harmonic regression, Fourier terms, TBATS, MSTL, Prophet, intermittent demand, Croston's method, INGARCH, global models, gradient boosting, LightGBM, recursive vs direct forecasting, M5 competition, DeepAR, N-BEATS, temporal fusion transformer, foundation models, Chronos, TimesFM, zero-shot forecasting, probabilistic forecasting, quantile loss, pinball loss, CRPS, PIT histogram, conformal prediction, hierarchical forecasting, forecast reconciliation, MinT, coherent forecasts, forecast monitoring, statsmodels, fable, KFAS, bsts, sktime, nixtla

What's inside

  1. Applied Time Series, Part 3: Advanced Topics — State Space, Regimes, and Modern Forecasting Practice
  2. Chapter 1: The State-Space View: Kalman Filtering, Smoothing, and Signal Extraction
  3. Chapter 2: Structural Time Series, Time-Varying Parameters, and Nowcasting
  4. Chapter 3: Intervention Analysis and Causal Impact: Counterfactual Forecasting
  5. Chapter 4: Regime Switching: Threshold and Markov-Switching Models
  6. Chapter 5: Multiple Seasonality and High-Frequency Forecasting
  7. Chapter 6: Counts, Intermittent Demand, and Non-Gaussian Series
  8. Chapter 7: Machine Learning for Forecasting: Features, Global Models, and the M5 Lessons
  9. Chapter 8: Deep Learning and Foundation Models: An Honest Assessment
  10. Chapter 9: Probabilistic Forecasting: Quantiles, Scoring Rules, and Conformal Prediction
  11. Chapter 10: Hierarchical Forecasting, Reconciliation, and the Production Capstone

Read a free sample chapter