Mistakes We Made Building Automated Trading Bots
The most useful mistakes from our trading-bot project: overreacting to losses, weak evidence, custom ML complexity and confusing plausible logic with demonstrated edge.
Changing too much after losses
If every losing sequence triggers new filters, thresholds or exits, the experiment never accumulates comparable evidence. We learned to freeze versions and create a new branch for material changes.
Plausible logic is not enough
Orderflow and ML ideas can sound mechanically convincing while failing after costs and forward execution. The hypothesis must survive measurement.
Complexity creates more failure surfaces
Custom adaptive policies introduced reward, exposure and state-management problems in addition to the original signal problem. More intelligence in one layer can create fragility elsewhere.
Preserve the failures
Deleting weak versions destroys the most useful research record. V4, V4.4, V5.1 and V6 remain distinguishable so later comparisons have context.
How we use this lesson now
We keep experiment versions separate, publish unfavorable PAPER outcomes, record net closed-trade metrics and avoid converting an early result into a profitability claim. The purpose of the record is comparison and diagnosis.