Model 002 · in the lab

The chart is a picture

Model 001 was a formula: measure the wave, ride the wave. It traded real money in public, and the market beat it — the record is on its page. Model 002 starts from a different observation: no human trader computes a time series. They look at the chart. A chart is a picture, and reading pictures is something machines got genuinely good at. So we stopped feeding our model numbers and started showing it images — the same shapes you see when you open a ticker and something in your gut says up or down.

The hard part was not the model

The hard part was refusing to be fooled by it. A trading model can ace every test you give it and still know nothing — because it quietly memorized the past instead of learning anything about the future. It can memorize the era it trained on. It can memorize the exact minutes it was sampled at. Backtests reward all of this, which is why the internet is full of strategies that never survive contact with Monday.

So we built a gauntlet instead of a backtest. Every candidate trains on one stretch of history and is judged only on days it has never seen, sampled independently, with tests specifically designed to catch each way a model can cheat. Then we try to kill the result again from another angle.

Five versions of Model 002 died in that gauntlet. Some of them looked spectacular first — one scored numbers that would make a hedge fund blush, right up until we moved the measuring points thirty minutes and watched it fall to a coin flip. We wrote every failure down. Killing your own results is the entire job; the market does it for you otherwise, at worse prices.

The one that would not die

The sixth build is different, and we are keeping how it works to ourselves. What we will say: it is built around an instinct any veteran chart-reader would nod at — old as trading floors — and it reads a chart the way you would tell a beginner to: the big picture first, then the last few hours, in that order. When we finally gave the machine that discipline, the coin flip broke.

On strictly unseen data — future days it could not have memorized, collected independently, net of trading fees — it kept its edge where every predecessor collapsed to chance. It is selective about when it speaks, and it is honest about one more thing: at longer horizons it knows nothing, and says so.

What happens next

The same thing that happened to Model 001, because it is the only test that counts: it goes live in public. Every call it makes will be logged the moment it makes it and scored against what the market actually did — a running record you can check, before and above any real money. A good backtest earns a model exactly one thing here: the right to be embarrassed in front of everyone.

The usual honesty, stated plainly: past performance — including ours on unseen test data — guarantees nothing about the future. Nothing here is financial advice. The record will say what the model is. Watch it.