Think Bayes Bayesian Statistics in Python (Allen B. Downey)（Z-Library）

【One-Line Pitch】 A hands-on introduction to Bayesian statistics for people who can already program: you'll build intuition by writing Python code for real problems instead of grinding through continuous mathematics. Best for developers, data folks, and self-learners who want a working grasp of priors, posteriors, and updating. 【Book Arc】 - **Opening (~0%–10%)**: Sets the philosophy — solve statistics with code, not formulas — and frames modeling as an explicit, first-class step. Introduces the idea that modeling errors usually dwarf numerical errors, so an approximate answer to a good model beats an exact answer to a bad one. - **Early (~10%–32%)**: Builds the foundation through counting-based probability, conditional probability, and Bayes's theorem, then moves into distributions and estimating proportions. Works through classic toy problems (coins, dice, cookies, the Linda problem) to make priors, likelihoods, and posteriors concrete. - **Middle (~32%–55%)**: Extends the toolkit to odds and addends, sums and mixtures of distributions, and estimation problems like the Train/German Tank problem. Here the book confronts the practical question of choosing priors — informative vs. uninformative — and argues for informative priors. - **Late (~55%–85%)**: Applies the machinery to richer models: Poisson processes (e.g., predicting soccer goals), decision analysis, and hypothesis testing. These chapters show how the same update loop scales to real-world estimation, prediction, and decision problems. - **Ending (~85%–100%)**: Closes with more advanced topics such as mixture models and Bayesian bandits, tying the computational approach back to applied decision-making. (Excerpts cover the table of contents and early chapters in most detail; later chapters are only sketched.) 【Key Takeaways】 - **Code replaces calculus as the primary tool** (Opening): The book's central bet is that discrete distributions and Python updates teach Bayesian reasoning faster than continuous math for programmers. - **Modeling is an explicit, unavoidable step** (Opening): Every chapter starts from a real problem, and you must decide what to include and abstract away — the book treats modeling error as the dominant source of error. - **Bayes's theorem has a diachronic reading** (Early): Prior, likelihood, and posterior describe how belief about a hypothesis changes as data arrives, not just a static formula. - **Priors are a deliberate choice, not a neutral default** (Middle): The author favors informative priors, arguing that with lots of data the choice barely matters and with little data background information helps a lot. - **Odds and likelihood ratios make updates intuitive** (Middle): Bayes's rule as posterior odds = prior odds × likelihood ratio lets you reason about updates by hand or in your head. - **The same update loop scales to real problems** (Middle/Late): From cookies to the German Tank problem to Poisson goal-scoring models, the pattern of prior × likelihood → normalize recurs. - **Decision analysis and bandits connect inference to action** (Late): The book moves from estimating parameters to choosing actions under uncertainty, including Bayesian bandits. - **You must run the code to learn it** (Early): The book is built from Jupyter notebooks, and the author explicitly says reading alone won't get you far — work the exercises. 【Reading Tips】 - **Deep-read the early chapters (probability, distributions, estimation)**: These establish the update pattern everything else reuses; skimming them will make later chapters feel like magic. - **Skim the historical asides** (e.g., the German Tank problem) on a first pass, then return if you want the applied-statistics context. - **Run every notebook and attempt the exercises** before moving on — the book is designed as a course, and the exercises (Linda problem, Elvis twin problem, dice games) are where intuition actually forms. - **Watch the floating-point and normalization details**: The author notes normalizing once at the end is efficient but can cause numerical issues, a practical gotcha worth understanding. - **Treat the prior-choice discussion as a decision point**: When you hit informative vs. uninformative priors, pause and decide what you'd do in your own problem. 【Coverage Limits】 This guide is based on stratified excerpts that cover the preface, table of contents, and early-to-middle chapters in detail; later chapters (decision analysis, bandits, and beyond) are only partially represented, so specifics there are inferred from headings and brief mentions.