Our own neural model reads the market we record — the full order book, the trade tape and years of history — and decides: buy, sell or hold. Watch every exam it takes, live. When a version passes, you receive its decisions. The model stays with us.
On its latest exam the model chose to buy 8%, sell 12% and hold 80% of the time. The animation above is an illustration of the four steps, not a live decision.
Every dot is one complete exam of a new version — hover it for the details. The blue line is the best result so far. Below zero means the version lost money after fees — those versions never leave the lab.
Results are the mean weekly return of a $100 account after exchange fees. Early versions lost up to 22 % a week; results below −8 % are drawn faded at the bottom edge.
123 weekly accounts of $100 — 41 held-out weeks × BTC, ETH, SOL.
Each square is one week of one market in the latest exam; lit squares ended in profit.
Each tower is a generation of the model, each cube an exam. A new generation starts when we change what the model sees or how it is taught. Hover or tap a tower.
Best exam so far: G7 v9, +0.75% per week after fees. Profitable exams: 3.
One decision per market whenever the model acts; silence is a decision too.