The day I questioned a 100% win rate

Where this starts

I run a tool I built myself that watches gamma data in the options market. When certain conditions line up it fires an alert, and later it grades that alert automatically — was it right or wrong.

One day I opened the grades.

Signal A   55 of 55 correct — 100%
Signal B   31 of 31 correct — 100%
Signal C   42 of 42 correct — 100%
Signal D   12 of 12 correct — 100%

Not a single miss.

Was I happy? No. I was scared.

Why that was frightening

No signal is right 55 times in a row in this market. If it is, it isn't a signal — it's a definition.

And there was a second odd thing. Other signals in the same tool were doing terribly.

Signal	Direction	Win rate
A	Down	100%
B	Down	100%
C	Down	100%
D	Down	100%
E	Up	8.6%
F	Up	13.0%
G	Up	26.8%

See it?

Every down signal perfect. Every up signal under 27%.

The split wasn't by quality. It was by direction. That is not a story about markets. That is a story about code.

The culprit

My database has a table that stores index prices. The problem: the S&P 500, SPY, and QQQ were all sitting in the same column.

S&P 500  →  7,719 ~ 7,816
SPY      →    769 ~   779
QQQ      →    716 ~   734

The grading logic read like this:

"If the price drops 20 points or more within a set window after the alert, a down signal counts as correct."

But the query that fetched "price" had no ticker filter.

So when it looked for the low, it found QQQ's 716.

716 − 7,800 = −7,084 points

A 7,084 point drop. That clears a 20 point threshold every single time.

Every down signal was being marked correct, unconditionally.

The high, on the other hand, happened to come back as the S&P's own value — which is why only the up signals showed realistic (low) numbers.

After the fix

One ticker filter, then I re-scored 329 alerts.

Signal	Before	Re-scored
A	100%	15.9%
B	100%	7.4%
C	100%	21.9%
D	100%	36.4%
E (up)	8.6%	8.6%
F (up)	13.0%	13.0%

The down signals collapsed. The up signals didn't move. That's how I knew the diagnosis was right.

And this is the real result: not one of the 11 signals reached my promotion threshold of 70%. The best was 36.6%.

Then it got worse

While checking the same code I found a different kind of 100%.

Signal X   1,599 correct / 0 wrong
Signal Y      64 correct / 0 wrong

This time it wasn't the ticker filter. The grading logic said:

move of 0.3% or more  →  correct
otherwise             →  undetermined

There was no "wrong" outcome at all.

Worse, it measured the move as an absolute value. Up or down, 0.3% counted as a hit. A directional signal, graded without looking at direction.

That isn't a win rate. It's an identity. It was built so it couldn't fail.

Over 1,599 records, the information content was zero. I deleted that signal.

What I took from it

1. A number that good is a bug, not a skill

Zero misses in 55 tries does not exist in this market. When a result is astonishing, open the code before you celebrate.

2. When two opposite directions score at opposite extremes, look at the code

Up 8.6% / down 100%. Markets are not that generous. Asymmetry like that is a symptom, not a signal.

3. Write the "wrong" condition first

When you design a test, write down what would make this signal wrong before anything else. If you can't write it, the thing isn't testable.

A test that can't fail isn't a test.

4. A table holding multiple tickers needs a filter on every read

Miss it once and the result inverts completely. And it doesn't throw an error. It quietly returns a wrong answer.

Working with Billie

Billie is an AI I built for my own trading. It isn't a general chatbot — it is wired into my observation logs and my trading database, and it remembers the rules I've written over the past several months.

I would not have found these bugs alone. The codebase is tens of thousands of lines, and the grading logic is not somewhere I look every day.

But Billie didn't find them on its own either.

It started with my question.

"What exactly are these four 100% signals measuring?"

Billie did not open with "that's a bug." I felt something was off and asked, and then we opened the data together.

Writing and coding with AI is something everyone does now. That's a baseline, not an edge.

What doesn't copy is being the person who says "this number looks wrong." That question comes from watching the market every day.

Where it left me

I turned off every alert in the tool.

None of the 11 signals cleared the bar, so there was nothing to leave running. I'll bring them back one at a time, only the ones that pass a real test.

It feels empty. Those alerts had been firing every day for months.

But silence beats a wrong alert. Especially when I think about what would have happened if I had traded on a 100% signal.

This is a record of how I validated a tool I built. It is not a recommendation to use any strategy or trade any security. Your decisions and their outcomes are your own.