I Scored Every Trade's Confidence. The Number Predicted Nothing.

Every trade my bot opened carried a confidence number. Eighty-five meant "decent setup." Ninety-plus meant "I'd bet the account." I built the whole thing assuming that number meant something — that when I felt surer, I won more. It's the most natural belief in trading, and mine was wrong in the most boring way possible.

I pulled the closed trades and split them by their confidence tag. Here is the entire argument in one small table.

Same win rate. Opposite P&L.

Read those win-rate cells again. Forty-one percent. Forty-one percent. Identical to the point. The confidence score — the thing my whole entry logic was built around — did not move the hit rate one single percentage point. The trades I graded a 97 lost exactly as often as the ones I graded an 86.

What did change was the money. The "decent" bucket bled −$3.59; the "bet the account" bucket scraped out +$0.27. Same frequency of losing, wildly different cost of losing. That gap is the whole lesson: confidence didn't predict whether I'd lose. It predicted, weakly, how much the losers would take when they came.

Why a win rate can lie to your face

Win rate is a headcount. It tells you how many trades finished green and nothing about their size. Two strategies can share a 41% win rate while one prints money and the other dies, because the only things that actually matter — average win, average loss — are invisible in that single number.

My lower-confidence trades weren't losing more often. They were losing deeper. On the setups I felt shakier about, I let the stop drift, I gave the trade "a little more room," I averaged down to prove myself right. The result was a fatter average loss on the exact same hit rate. The 90+ trades got the same 41% wins but I cut their losers faster, so the damage was smaller. Discipline, not conviction, was doing the work.

This is the same disease I keep finding in my data. One microcap in a big GPU backtest came back at a 96% win rate with a 3.47 profit factor — and hid a −14.2% worst-case excursion propped up by averaging down 64.8% of the time. Almost never lost. When it did, it took a chunk that needed +16.6% to repair. High win rate, ugly shape. The headcount looked perfect and the P&L told a different story.

The whole account said it too

Zoom out to all five of my paper systems: 81 trades, a 15% net win rate, and a combined result of −$0.567. A system that loses 85% of its trades landed almost exactly at flat, because the rare winners were large and the frequent losers were tiny. That's expectancy doing its job. It is the opposite of the confidence story and the same math — the number of wins told me almost nothing about the money.

Confidence felt like information. It was emotion wearing a number. Once I bucketed it, the correlation with winning was flat, and I'd been sizing on noise.

What I changed

1. Stopped sizing on conviction. If the confidence tag doesn't lift the win rate or the expectancy, putting more money behind a "97" just risks more on a coin flip. Size comes off risk now, not off feelings.
2. Grade the loss, not the setup. The number I care about at entry is the capped loss if I'm wrong, because that's the variable my data says actually drives the result.
3. Judge every bucket by expectancy, never by win rate alone. A 41% win rate is fine or fatal depending entirely on the size of the average win and loss — so I compute expectancy before I trust any hit rate.
4. Cut the shaky trades faster, not slower. The "decent" bucket lost because I gave weak trades more rope. Now the less sure I am, the tighter the leash, not looser.

→ Compute your real expectancy · → Break-even win rate · → R-multiple calculator