Building an Honest AI Trading Agent

By Sérgio José Rodrigues Saraiva

Modern markets are close to efficient. Point a machine-learning model at price history, evaluate it honestly, and the up-or-down call slides back toward a coin flip.

Most books about AI and trading never report that, because most never measure it. This one measures everything - and then asks the question that actually matters: what is left to trade once you accept it?

I built a production, multi-asset autonomous trading agent - designed it, deployed it against real broker APIs, and spent the better part of a year trying to break it. Roughly forty directional experiments, each held to leak-free data handling, walk-forward validation, calibration, and statistical tests strict enough to catch a lucky backtest before it catches you. This book is the whole build, end to end, and the whole record of what survived that scrutiny and what did not.

It is also the answer. Something does survive - not a comforting "buy signal," but a different axis of the market entirely, one you can measure, size, and compound with the same rigor used to find it. Getting you there, honestly, is what this book is for.

Inside, you will build: Multi-timeframe feature engineering and leak-free labeling (the triple-barrier method) across five asset classes - forex, crypto, commodities, indices, and stocksPer-timeframe gradient-boosted models plus a stacking meta-learner, validated by walk-forward optimization across nine rolling windowsAn Observe → Filter → Size → Execute → Monitor → Reflect trading agent, wired to real broker APIsA twelve-layer circuit-breaker stack for volatility, drawdown, and portfolio-heat protection, with every position sized by the signature formula: R = P K ΩThe statistical machinery - Deflated Sharpe Ratio, Probability of Backtest Overfitting, forward-AUC - that tells a real edge from a lucky backtestForty documented experiments in the systematic search for direction, and the one axis that came through them intact

Every central claim is backed by a runnable, open companion codebase, so you can reproduce the honest numbers yourself, then point the same leak-free pipeline and statistical toolkit at your own ideas.

This is the book that hands you the instruments, not just the promise: a validated edge to build on, and the discipline to tell your next real discovery from your next false one.

• Format: Paperback
• Language: English
• ISBN: 9893402212
• ISBN13: 9789893402214
• Release Date: July 2026
• Publisher: Sergio Saraiva
• Length: 538 Pages
• Weight: 1.41 lbs.
• Dimensions: 1.3" x 6.0" x 9.0"