In the previous two stories we dug into the common beginner mistakes, and then into the “smart stop-loss” an AI promised me but never actually built. Today’s war story is uglier, because the enemy wasn’t a bug. The enemy was a number that looked like success.

The number that lied

I had a bot with an 80% win rate. Eight out of ten trades closed green. On paper it looked like a money printer, and for a while I bragged about it like one.

The account balance disagreed. It kept drifting down.

How does a bot win 80% of the time and still lose money? Here’s the math nobody puts on the marketing banner:

80% win rate at a risk/reward of 1 : 0.3 means you risk $1 to make $0.30. 8 wins × $0.30 = +$2.40 2 losses × $1.00 = −$2.00 …minus fees and funding, and you’re underwater.

The win rate was real. It was also completely useless on its own. A high win rate with a bad risk/reward ratio is the most expensive vanity metric in trading.

Why AI makes this trap worse

When you ask an AI assistant to “optimize my bot,” it will happily tune parameters until the win rate looks beautiful — because win rate is easy to maximize. Tighten take-profit, widen stop-loss, and watch the green percentage climb while your expectancy quietly goes negative. The AI did exactly what I asked. I asked the wrong question.

What I built instead

I stopped optimizing for win rate and started optimizing for expectancy:

expectancy = (win_rate × avg_win) − (loss_rate × avg_loss)

If that number isn’t positive, nothing else matters — not your win rate, not your “AI-powered” anything.

Then I did the thing that scares most people selling courses: I made the metric impossible to fake. Our live paper-traded bot publishes its real numbers to a dashboard that updates itself. Today it runs around a 56% win rate — lower than that shiny 80%, and profitable, because the risk/reward is positive. I would rather show you an honest 56% than a flattering 80% that loses money.

The lesson

• Win rate alone tells you almost nothing.
• A high win rate + bad R:R = a slow-motion loss.
• Optimize for expectancy, then look at drawdown, then everything else.
• If a bot (or a course) only advertises win rate, ask what the risk/reward and the drawdown are. The silence is the answer.

I teach exactly how to build, measure and stress-test this yourself — with AI doing the coding and you doing the thinking — in the course. The first module is free, and the bot’s live numbers are public, drift and all:

👉 https://nexus-bot.pro/ — build your own bot with Claude Code, no black box. 📊 Live results (the honest 56%, not a screenshot): https://nexus-bot.pro/proof/rvv/

Next war story: the day my backtest showed 400% returns and I almost believed it. (Spoiler: look-ahead bias.)