Descriptions:
This video from the Trade Tactics channel walks through the development and rigorous testing of a trading bot built around the Kalman filter — a mathematical technique originally created by Rudolf Kalman for NASA’s Apollo program in 1960. The creator applies a three-stage Kalman filter (tracking price position, velocity, and acceleration) to strip noise from trading signals, using Claude Code to automate overnight strategy generation and parameter optimization.
A central focus is the look-ahead bias problem: the creator discovers that an apparently stellar strategy (2.35% drawdown, near-perfect bootstrap results) was invalid because Claude had inadvertently built in access to future candlestick data. The video clearly explains how this manifests as repainting in live trading and how to guard against it in Python by explicitly banning forward-looking calculations.
The testing methodology is notably rigorous for a YouTube trading channel: strategies are evaluated across three distinct data splits — a training set Claude can tune on, a preliminary out-of-sample set, and a fully held-out final set Claude never touches — plus Monte Carlo bootstrapping for robustness. Surviving candidates show Sharpe ratios around 0.7–0.8 and drawdowns in the 24–32% range. The video also briefly covers the “super smoother” filter as an alternative signal-smoothing approach.
📺 Source: Trade Tactics · Published March 21, 2026
🏷️ Format: Hands On Build
