This book is written for people who want to build systems that learn from time: thoroughly, from the autocorrelation function up to the planning agent, without first acquiring a specialist degree in time-series analysis. If you have seen linear algebra, probability, and a first course in machine learning, you have everything you need to start at Chapter 1 and finish at the capstone. It serves senior undergraduates and graduate students as a course text, researchers as a unified reference across traditions that rarely sit on one shelf, and practitioners as a working manual for shipping temporal systems that survive contact with production.
You Are in the Right Place If...
- You are the forecasting engineer. You already fit ARIMA and gradient-boosted models on business series, and you want to understand the deep and foundation models that keep appearing in benchmarks, plus the uncertainty and drift machinery that decides whether a forecast is trustworthy. Parts III, V, and VIII are written for exactly your next step.
- You are the machine learning researcher moving into sequences. You know training loops, attention, and embeddings from vision or language, but temporal structure (stationarity, leakage, irregular sampling, the classical baselines) is unfamiliar terrain. Parts I and II give you the time-specific foundations your background skipped, and the bridge boxes connect them to machinery you already trust.
- You are the reinforcement-learning practitioner. You think in policies and value functions and want to see decision making as the natural continuation of sequence modeling rather than a separate world. Part VI builds RL on the filtering and state-space ideas of Part II, and shows the Decision Transformer treating the Markov decision process as just another sequence to predict.
- You are the instructor. You need a coherent two-semester sequence with measurable objectives, three-flavor exercises (theory, implementation, open-ended), running datasets for comparability across chapters, and a capstone that turns knowledge into a portfolio artifact. Parts I through VII map onto a standard sequence; Parts VIII and IX work as a project capstone or assigned reading.
- You are a self-taught learner or a student. The book is self-contained, sequenced for cumulative reading, and defines every term it uses, so a determined reader can travel the whole arc without a course around it.
What You Need Before Page One
Four modest prerequisites. Linear algebra: vectors, matrices, eigenvalues, and comfort with a transformation written as a matrix. Probability: distributions, expectation, conditional probability, and the idea of a random process. Calculus: derivatives, gradients, and a little integration. Introductory machine learning: what a loss function is, how gradient descent works, the meaning of training and test splits. You do not need to be expert in any of these; the code favors clarity over cleverness and the math appears where it earns its keep. Appendices A through D refresh every piece the chapters rely on, exactly when intuition is no longer enough: unified notation and mathematical foundations (A), probability and statistics (B), optimization and deep learning (C), and PyTorch for temporal work (D). When a chapter assumes something, it tells you which appendix to consult.
Hardware is not a barrier. Parts I, II, and V run comfortably on any laptop CPU. The deep-learning, foundation-model, and reinforcement-learning chapters benefit from a GPU, but the examples are sized so that a modest consumer card or a free cloud notebook suffices; nothing in the book requires a data-center budget.
What You Do Not Need
You do not need prior time-series experience: the book assumes none and starts from what a time series is. You do not need a statistics or control background: the classical, neural, and decision-making traditions are introduced from first principles and, more usefully, reconciled, so the Kalman filter, the recurrent network, and the belief-state agent end up speaking one language. And you do not need allegiance to a particular framework: the book teaches concepts first and reaches for the right mainstream tool for each job, statsmodels and the Nixtla stack early, PyTorch, Lightning, and the foundation-model ecosystem later, Gymnasium and the standard RL libraries for the decision chapters.
What This Book Is Not
It is not a theory monograph; proofs are sketched only when they change how you build. It is not a forecasting-only cookbook; prediction is the first third of the story, not the whole of it. And it is not an exhaustive survey of the literature; each chapter curates a short annotated bibliography rather than citing everything, aiming for the understanding that survives the next library release.
If that matches your situation, continue to the next page for a guided preview of what the chapters look like inside, or open the Table of Contents to choose a reading path.