CRYPTO SELF-LEARNING FORECAST

How to read the ladder. Each rung is a wall-clock time, not an abstract horizon: at 11:30 the rungs answer 11:35, 11:40 and 11:45. The gold pin is the model's point guess; the green band behind it is where price actually lands 95% (light) and 68% (solid) of the time. The pin sits almost exactly on the centre line, and that is the honest result, not a rendering bug. Only the 15-minute rung is scored below — the 5 and 10 minute rungs pro-rate the same drift by time (drift ∝ h) while the band grows as σ√h, so the shorter rungs are arithmetic on one forecast, not three independent ones.

What "self-learning" means here. Every minute the model makes a real point forecast — the +15 min rung — as current price plus a small drift from an online linear model. Its features are price momentum & mean-reversion plus live microstructure: order-flow imbalance (buy vs sell volume) and book imbalance (bid vs ask depth), streamed from Coinbase trades and the order book. Fifteen minutes later it looks up what actually happened, measures the error, and nudges its weights by gradient descent. It keeps score against the random-walk baseline (predict "no change") and remembers everything across sessions.

Watch the weight bars. They show what the model has actually learned to rely on. My honest bet: the order-flow weights stay near zero at a 15-minute horizon — those signals are real but decay in seconds, so they're near-useless this far out (they'd matter far more on a 1-minute forecast). If an order-flow weight grows and Skill turns positive, that's a genuine lead worth a proper backtest.

Set expectations honestly: at a 15-minute horizon price is near-unpredictable, so the most likely thing the model learns is to shrink its weights toward zero and forecast ≈ the current price — i.e. it discovers there's no edge. The scoreboard proves it either way: Skill > 0 and Hit-rate > 50% over many resolved predictions would be a real signal; hovering around 0 / 50% means random walk wins. It's an honest experiment and an ML demo, not a trading signal — no live trade without a proper backtest.