The book reads best front to back: it was sequenced so that every idea lands on ground prepared by an earlier chapter, and a recurring temporal thread keeps returning to the same handful of structures in ever more powerful forms. The Kalman filter of Chapter 7 comes back as the recurrent network of Chapter 10 and the structured state-space model of Chapter 13; the hidden Markov model returns as a neural latent-variable model in Chapter 17 and a belief-state agent in Chapter 23; the Markov decision process of Chapter 22 returns as a sequence model in Chapter 28. But the book was also built for readers with deadlines, and the dependencies between parts are explicit enough that several shorter paths work well. This page maps the dependencies first, then offers three tested reading paths.
How the Parts Depend on Each Other
Part I is the foundation for everything: its data types, leakage and backtesting discipline, and temporal-engineering conventions are assumed by every later part. Part II underpins Part III, because the classical models it teaches (autoregression, state-space filtering, spectral analysis) are the structures the deep architectures later learn from data, and the bridge boxes only pay off if you have met the classical side first. Part VI (sequential decision making) builds on Parts I through III: it assumes you can engineer temporal data, reason about a stochastic process, and read and train a neural sequence model before policies and value functions arrive. Table F5.1 summarizes the load-bearing edges.
| If you are heading for | Make sure you have read |
|---|---|
| Part II (classical forecasting) | Chapters 1 and 2 from Part I |
| Part III (temporal deep learning) | Part I, plus the relevant classical model in Part II (Chapter 7 before Chapter 10 and 13; Chapter 5 before Chapter 14) |
| Part V (uncertainty, online, adaptive) | At least one forecasting chapter from Part II or III |
| Part VI (sequential decision making) | Parts I through III, with Chapter 7 especially useful before Chapter 23 |
| The capstone project | At least one full pass through Parts I through III and one chapter of Part VI |
Table F5.1 The dependency edges that matter when skipping ahead.
Three kinds of material sit outside the dependency graph and can be visited at any time: the appendices backstop the mathematics, probability, optimization, and PyTorch on demand; the dataset and benchmark catalog in Appendix E is useful from the first project onward; and the three running datasets (a financial returns and volatility series, an irregularly sampled clinical series, and an industrial sensor stream) recur across the parts so that a method met in one chapter can be compared against another method met fifty pages later.
Path 1: The Practitioner
You want to ship a working temporal model soon and deepen later. Skim the theory for intuition rather than reading every derivation, and instead spend your time running the code and labs at each section's end. Read Chapters 1 and 2 carefully, because leakage and backtesting mistakes are the ones that hurt in production. Then go straight to the model family your problem needs: Part II for a strong statistical baseline (ARIMA and exponential smoothing in Chapter 5 will often be hard to beat), Part III for deep forecasting (Chapter 14) or a foundation model you can prompt zero-shot (Chapter 15), Part V for calibrated intervals and drift handling, and Part VI if your problem is really about acting rather than predicting. Keep the library-shortcut callouts open as a cookbook, and revisit the theory when a model misbehaves and you need to know why.
Path 2: The Researcher
You want depth, current methods, and the trail into the literature. Read for the full depth, including the derivations and the bridge boxes that connect classical and neural methods, and pay particular attention to the research-frontier callouts and the chapter bibliographies, which are curated as entry points into the 2024 to 2026 literature. Read Part II in full for notation and for the structures that return later, then read Parts III through VII completely and in order; the temporal thread is most rewarding when followed unbroken. Appendix E gives you the benchmark landscape for positioning experiments, and Appendix F covers reproducibility and compute so that your results survive contact with a reviewer.
Path 3: Course and Self-Study
You are teaching from the book or learning at your own pace. Parts I through VII map onto a standard two-semester sequence: a first semester covering Parts I, II, and III (foundations, classical forecasting, and temporal deep learning), and a second semester covering Parts IV, V, VI, and VII (representation learning, uncertainty and adaptation, sequential decision making, and intelligent temporal systems). Part VIII (trustworthy and deployed temporal AI) and Part IX (applications and future directions) work well as a project capstone or as assigned reading, depending on course emphasis. Do the exercises, which are split into theory, implementation, and open-ended problems: the theory ones before moving on, the implementation ones for any chapter touching your interests, and the open-ended ones whenever a claim surprises you. Start sketching a capstone while reading Part III so that the later parts can feed it.
How to Run the Code
Every code example is Python and runs against the modern temporal toolbox. Classical methods use statsmodels, sktime, pmdarima, Prophet, and the Nixtla family (statsforecast, neuralforecast, mlforecast); deep forecasting uses PyTorch, PyTorch Forecasting, Lightning, GluonTS, and Darts; foundation models come through Hugging Face (Chronos, Lag-Llama, Moirai, TimesFM, MOMENT); online learning uses river; and sequential decision making uses Gymnasium, Stable-Baselines3, CleanRL, and d3rlpy. Appendix D walks through the environment setup and a PyTorch refresher for temporal data, with JAX notes for readers who prefer it. Each section that teaches a concept from scratch also gives a library-shortcut callout that solves the same task in a few lines and states the line-count reduction, so you always see both the mechanism and the production-ready path. From the table of contents you can reach any chapter, appendix, or the capstone directly.
The Callout System and the Per-Chapter Scaffold
Each chapter follows the same scaffold so that teaching and self-study stay predictable: learning objectives open the chapter, a motivation and running example set the stakes, the core exposition carries the theory with derivations and figures, fully worked examples solve representative problems, a classical-to-neural bridge box names the structure that returns elsewhere in the book, a case study applies one running dataset end to end, a pitfalls section flags leakage and look-ahead traps, exercises close the technical content, and a further-reading section points into current work. Inside the prose, the callout system carries the asides: practical-example boxes ground a method in a finance, healthcare, or industrial story; library-shortcut boxes give the few-line production path; research-frontier boxes mark open questions; and note, warning, and definition boxes handle the rest. Run the code, use the cross-reference links generously so that a forgotten prerequisite is one click away, and when a chapter ends glance at its What's Next section even if you are about to jump elsewhere, because that is the narrative thread that keeps the nine parts feeling like one book.