V5.1: Lessons From Our Adaptive ML Trading Experiment

V5.1 was an attempt to let a trading engine adapt more of its behavior from live market flow. The experiment produced useful failures: policy adaptation, reward design and portfolio behavior were harder problems than simply adding machine learning.

What we were trying to build

The pipeline consumed OKX market data, generated features, selected LONG, SHORT or FLAT behavior, sized positions, managed exits and used realized net outcomes as learning feedback. Multiple policies represented trend/momentum, reversal, breakout and no-trade behavior.

DATA→FEATURES→POLICY→PAPER TRADE→REWARD

What went wrong

OBSERVATIONLESSON

Behavior did not adapt reliably after lossesAn adaptive label does not guarantee useful adaptation; the feedback loop itself must be designed and tested.

Conflicting long/short losses appearedPer-trade intelligence is insufficient without portfolio-level exposure control.

Reward needed revisionThe objective must represent the behavior we actually want after costs and risk.

Different V5.1 branches behaved differentlyResults must stay attached to their exact policy/version rather than being merged into one ML narrative.

The result was not deleted

One historical V5.1 adaptive sample recorded 16 closed PAPER trades with 3 wins and 13 losses and negative realized net PnL. A separate production-policy sample behaved differently. These are small historical samples, but preserving both prevents us from presenting only the favorable branch.

Why V6 followed

Instead of continuing to expand a custom ML engine, we moved the next experiment to Freqtrade/FreqAI. That gave us a more mature trading framework while keeping the research question focused on model and policy behavior.

How we built V6 →

Related research

Why AI bots fail →FreqAI training →