What the model thinks, hour by hour.

Why does this number look different from Track Record?

They measure two different things on purpose. Trade Win Rate (on the Track Record page) is whether a real, executed trade closed in profit — it can win on a tiny locked-in gain from the profit-lock mechanism even if the price keeps moving after the trade is already closed. Hourly Lean Accuracy (this page) is stricter and unrelated to whether a trade was even open: did the model's directional call for that hour match where the price actually was one full hour later, no matter what. A trade can close as a real win while that hour's lean still grades "wrong" here, and vice versa. Neither number is wrong — they're answering different questions.

We tried to build a real hourly-direction model. Here's what happened.

The cards above show the trading model's own confidence, repurposed into an hourly up/down lean — useful for transparency, but never validated as an actual forecaster. So we built one and tested it properly: a dedicated classifier per symbol, trained on 7–10 months of real H1 candles pulled directly from the broker (not synthetic data), same feature pipeline as the live trading model, evaluated on a strict chronological 80/20 split with zero lookahead.

An AUC near 0.50 means no real edge — the model is guessing. Across all five symbols, accuracy barely moves off (and in three cases falls below) simply always guessing the majority direction. This matches well-known finance research: an hour of price movement in a liquid market is close to a random walk, and technical indicators alone don't reliably predict it. We're not treating this as a trading signal or using it in the live bot — but since this whole site is built on showing real numbers instead of hiding bad ones, we're running it forward anyway, hour by hour, starting Sep 18, 2026. The section below is that live, ungraded-in-advance track record.