The first live bot: 83 percent wins, zero profit

The first rule that went live was the obvious one: at three and a half minutes into a 5-minute window, buy the favourite if it costs 70 to 90 cents, hold to settlement. It won 83 percent of its trades. It made nothing.

Why a high win rate is not an edge

An 83-cent favourite that wins 83 percent of the time returns exactly what you paid, minus the fee. The book is calibrated. The fee is not. The rule was a slow way of paying 0.9 cents a share to the exchange.

What broke, in order

1. Winnings sat unredeemed. Settled shares stay in the wallet until redeemed; the balance looked wrong and new orders were refused at a few hundred dollars of unclaimed wins.
2. An order that errored had filled anyway. A buy raised an exception on the wire, matched seven seconds later, and there was no position row. The fix was a background check of the fills endpoint after every exception.
3. Eleven percent of the labels were wrong. Outcomes derived from the last oracle read disagreed with the official settlement on 11.4 percent of windows. Every backtest run before that discovery was built on bad labels.
4. A 14 percent edge was the labels. A "latency edge" that survived two weeks of checking disappeared the day the official outcomes were backfilled. The fake edge came from the same feed that produced the bad labels; the circle closed on itself.
5. The stop-loss cost more than it saved. Every exit rule tested, including a learned one, lost to holding. A false stop turned a small win into a larger loss more often than a true stop rescued anything.

What it cost

Not much money; a first-day loss in the tens of dollars on a small bankroll. What it cost was three weeks of believing a rule that was breakeven by construction, and the lesson that a win rate above 80 percent means nothing until you subtract the price paid.

What it left behind

• A logger that records the official outcome by market id, not a derived one.
• A fill reconciler that runs after any exception.
• A daily redemption sweep.
• The habit of scoring every rule as win rate minus breakeven at the price paid, per trade, never pooled.