Part Overview
Every model in this book, from an ARIMA forecaster to a Decision Transformer, reasons about the same primitive: a quantity that changes over time. Part I is where that primitive is named and tamed. Before you can predict a market, monitor a patient, or control a machine, you have to understand why time turns ordinary machine learning into something harder, what the theoretical ceiling on prediction actually is, and how to organize temporal data so your evaluation does not quietly lie to you. These two chapters are short, but they set the standards the other eight parts are held to.
Chapter 1 makes the case that time is a first-class dimension of intelligence rather than just another feature column. It surveys temporal data across finance, healthcare, industry, and agents; separates the three things temporal systems do (predict, reason, and decide); and introduces a crucial idea that most courses skip: predictability is bounded. Entropy rate and the limits of forecasting tell you, before you fit anything, how much signal is even there. The chapter closes with a map of the whole field and a guide to reading the book, including the recurring temporal thread that links each classical method to its learned descendant.
Chapter 2 turns to the unglamorous craft that decides whether anything later works: temporal data engineering. Time stamps and time zones, resampling and aggregation, irregular and missing observations, windowing and lag construction, and above all leakage-free backtesting. The single most common way temporal AI projects fail is look-ahead bias, and this chapter builds the habits that prevent it. The three running datasets of the book (a financial returns series, an irregularly sampled clinical series, and an industrial sensor stream) are introduced here so that later parts can compare methods on shared, well-understood data. Read Part I with a notebook open; the conventions you internalize now are the ones you will rely on for the remaining 34 chapters.
Why time is a first-class dimension of intelligence, how prediction, reasoning, and decision making differ, and what entropy rate says about the limits of forecasting.
Time stamps, sampling, irregular and missing data, feature and lag construction, and the leakage-free backtesting that keeps every later result honest.
Where This Part Leads
With the vocabulary of temporal data and the discipline of leakage-free evaluation in hand, the book turns to its first family of models. Part II: Classical Forecasting and Time Series Analysis builds directly on these foundations: the stationarity and autocorrelation ideas sketched here become the statistical machinery of ARIMA and GARCH, and the state-space view introduced as a data structure becomes the Kalman filter, the first classical method the temporal thread will later return to in learned form.