This book is organized as a single arc in nine parts that moves from describing time to predicting it, from predicting it to representing and quantifying it, and finally from understanding it to acting on it. Thirty-six chapters carry you from "a time series is a sequence of timestamped observations" to "a sequence model can choose the next action that best reaches a goal." Seven appendices and a capstone project round out the journey. The full map lives in the Table of Contents; this page walks the arc so you know where each part fits and why the order matters. A single temporal thread runs through it: each classical idea returns in learned form, and the book is sequenced so that each return feels like recognition.
Part I: Foundations of Temporal AI
Part I establishes what makes intelligence temporal and why time resists the assumptions that make ordinary machine learning convenient. It asks what predictability even means, sets the limits that entropy and the structure of a process place on any forecaster, and then turns to the unglamorous craft that quietly decides whether a project succeeds: temporal data engineering. Here you learn to handle timestamps, resampling, and irregular observations, and to respect the leakage boundaries and backtesting protocols that separate honest evaluation from accidental time travel.
Part II: Classical Forecasting and Time Series Analysis
Part II builds the statistical bedrock: stationarity, autocorrelation, decomposition, and the frequency domain, then the univariate workhorses (AR, MA, ARIMA, exponential smoothing, Prophet) and their multivariate and econometric extensions (VAR, cointegration, GARCH volatility). It closes with state-space models and filtering, where the Kalman filter and the hidden Markov model first appear, and with anomaly and change-point detection. These methods are not nostalgia; they define the vocabulary and the baselines that every later chapter answers to.
Part III: Temporal Deep Learning
Part III is where most working temporal systems live today. It starts with neural sequence fundamentals and the gradient pathologies that motivated everything after, then builds recurrent networks (and the bridge that casts the RNN as a learned Kalman filter), temporal convolutions, attention and transformers, and the structured state-space and continuous-time models (S4, Mamba, Neural ODEs and CDEs) that unify the family. Two chapters then specialize the machinery for prediction: deep forecasting architectures (DeepAR, N-BEATS, TFT, PatchTST) and the temporal foundation models (Chronos, TimesFM, Moirai, MOMENT) that forecast unseen series zero-shot.
Part IV: Temporal Representation Learning
Part IV asks how to learn useful temporal features without labels. It covers self-supervised and contrastive methods for time series, masked modeling, and disentangled embeddings; generative temporal models from sequence autoencoders and VAEs to GANs and diffusion (TimeGrad, CSDI); and the modeling of events, point processes, and temporal graphs, where the hidden Markov model returns as a neural latent-variable model. Representations learned here feed every downstream task in the parts that follow.
Part V: Uncertainty, Online, and Adaptive Learning
Part V confronts the fact that a single number is rarely an honest forecast. It develops probabilistic and quantile forecasting, proper scoring and calibration, Bayesian and ensemble methods, and conformal prediction for time series. It then turns to the world that will not stay still: online and continual learning under concept drift, catastrophic forgetting and replay, and the adaptive systems that detect drift in production and decide when to retrain.
Part VI: Sequential Decision Making
Part VI, the largest of the decision-oriented parts, turns prediction into action. It builds Markov decision processes and dynamic programming, partial observability and POMDPs (where filtering returns as belief tracking), bandits and the foundations of reinforcement learning, deep and advanced RL, and optimal control with imitation learning. Two capstone chapters then close the temporal thread: sequences as decision making, where the Decision Transformer casts the MDP as a sequence model, and world models and planning in latent space.
Part VII: Building Intelligent Temporal Systems
Part VII assembles the pieces into systems that reason and act over time. It covers temporal reasoning and causality (temporal logic, causal discovery, treatment effects over time, counterfactuals), temporal AI agents with memory, planning, and tool use, and spatio-temporal intelligence for mobility, traffic, video, and dynamic graphs.
Part VIII: Trustworthy and Deployed Temporal AI
Part VIII addresses what serious practitioners cannot skip. It treats interpretability for temporal models, robustness and adversarial attacks, fairness, privacy, and federated learning, and the full deployment story: streaming inference, monitoring and drift alarms, feature stores, cost and scaling, and incident response. This is the part that keeps a temporal system honest after the demo ends.
Part IX: Applications and Future Directions
Part IX grounds everything in practice and then looks ahead. It walks industrial applications across finance, healthcare, manufacturing, energy, robotics, and scientific discovery, drawing on the three datasets that recur throughout the book, before closing with the open frontier: unified sequence models across modalities, foundation agents, scaling laws, and the path toward general temporal intelligence.
Beyond the Chapters
Seven appendices support the main text: mathematical foundations and a unified notation table that reconciles the clashing symbols of the statistics, machine-learning, and control traditions (A), refreshers in probability and statistics (B) and optimization and deep learning (C), a PyTorch-for-temporal guide with JAX notes (D), a datasets and benchmarks catalog (E), a reproducibility and compute guide (F), and worked solutions to selected exercises (G). The capstone project then asks you to put the whole arc to work in one end-to-end system: classical baselines, a deep or foundation forecaster, calibrated uncertainty, a decision policy, and an honest evaluation under drift.
One thread runs through all nine parts: ideas introduced classically return learned. The Kalman filter becomes the RNN and then the state-space model; ARIMA becomes linear attention; the hidden Markov model becomes a neural latent-variable model and a belief-state agent; the MDP becomes a sequence model. The next page describes who this journey is designed for.