Why do most trading bots fail?

The mechanism is rarely deliberate fraud. It happens through ordinary iteration — adjust a threshold, the backtest improves, keep the change — and after enough of those the system is shaped entirely by the noise in one sample. The resulting equity curve is genuinely beautiful and genuinely meaningless.

The second cause is hidden risk. Martingale and grid recovery produce smooth results for long periods by deferring losses rather than taking them, and the eventual sequence that the account cannot fund arrives without warning.

The third is cost blindness. Strategies tested with a fixed narrow spread and no slippage are profitable in a market that does not exist. On gold, where spreads widen exactly when signals cluster, this alone can flip the sign of the result.

What survival looks like

The defences are unglamorous: hold data back and test on it once, prefer fewer parameters, model costs pessimistically, and publish what failed alongside what worked. We tested 34 strategies and ship 4 — the other 24 and their reasons are public, because a vendor showing only their winners has told you nothing about their process.

We also removed a feature that improved in-sample results. Scaling reward-to-risk by market regime looked better on the fitted period and turned out to be fitting noise; a single 2.0 ratio everywhere made the system better by making it simpler.