The preface sets up why this book exists, what it covers, and how to read it. It introduces the running notation that the rest of the book uses without repeating, and it makes the prerequisites explicit so nobody runs into Chapter 6 expecting a probability primer.
"All we have to decide is what to do with the time that is given to us." — Gandalf, The Lord of the Rings: The Fellowship of the Ring (2001)
There is no shortage of deciding-what-to-do in quantitative finance. Every model trained, every position sized, every alert investigated is a small decision under uncertainty. What this book argues — and what each chapter will try to make concrete — is that the workflow for those decisions has finally matured enough to be written down end-to-end.
I have spent the last few years assembling that workflow piece by piece for my own work: a forecaster I could defend to a risk officer, a policy I could explain to a portfolio manager, a latent-state model I could intervene on for a research question, an agent that could glue the rest together when the deadline was tight. Each piece exists in excellent isolated treatments — there are textbooks on time-series forecasting, on portfolio choice, on state-space modelling, and now on agentic AI. What is harder to find is a single source that takes one problem all the way through, with the same notation, the same data, and the same code. This book is my attempt at that source.
A practical, end-to-end walkthrough of modern financial forecasting and decision systems, organised so each chapter feeds the next:
The aim throughout is to keep theory close to code. Every chapter
pairs the mathematical statement of an idea with a working snippet,
usually in Polars for data and
PyTorch for models. Longer runs live as
end-to-end scripts under codes/; the README there maps each script
to the section it backs.
It is not a substitute for a serious treatment of probability, optimisation, or asset pricing. Where those subjects are needed I quote the result and point to the standard reference. It is also not a tutorial on back-testing infrastructure or production deployment — those topics are adjacent and important, but each deserves its own book.
It is also not, deliberately, a survey. Every methods chapter has had to pick a small number of techniques to develop in depth rather than list everything in the area. Where a technique I think is genuinely useful did not make the cut, the relevant section says so and points to the literature.
The chapters are designed to be read in order, but each chapter index is self-contained so you can drop in to any topic and still see the plan. Chapter 1 is the entry point I would skim even if you read nothing else — it lays out the prediction → decision → dynamics → automation → synthesis pipeline, names the three reading paths through the book, and fixes the notation everything later uses.
If you are new to dynamic-factor or state-space thinking, the path I would recommend is modelling-first: Chapter 2 → Chapter 6 → Chapter 4 → Chapter 5, then Chapters 7–9 in order. If you already know the modelling and need the agentic plumbing, jump to Chapters 7 and 8. Detailed reading paths and the full notation table live in the Chapter 1 index.
This material grew out of teaching notes and a series of research projects on identifiable latent representations and recursive-utility reinforcement learning. I am grateful to the students who pushed back on early drafts, and to the open-source communities behind Polars, PyTorch, sktime, and Hugging Face for the tools that make this kind of book possible at all.