Continuous learning

What counts as evidence

Each coin contributes one return: the simple average of returns from its fully closed simulated positions, after modelled trading costs. Open positions and dip-entry strategies are excluded. Holder rewards are not included. A coin can contribute while another strategy still holds it; only closed positions enter its average.

That return is matched by coin to its first recorded Alpha score and factor values. Each screen measures Spearman rank correlation separately. The sample counts coins, not individual trades. Only coins the simulation traded are represented; these are historical associations, not an out-of-sample test.

With at least 20 pairs and non-constant values, a cycle moves the stored multiplier 25% toward the target (1 + correlation × 0.3), bounded to 0.70×–1.30×. Effective weight is base weight × stored multiplier. A positive correlation can still reduce a multiplier that is already above its target.

The server's background maintenance checks for a learning cycle; cycles are gated to at most once per 3 hours in the running process. The stored weights apply to future scores. Below the minimum, or with constant data, a factor keeps its previous multiplier. The tables show current evidence alongside stored weights, so a new correlation need not have been applied yet.

The 24h market-cap comparison is context only. Historical totals include retired screens; the diagram and tables show live screens. Trends compare only snapshots using the same trade-return target. Older simulated exits may include stale-price errors fixed on 2026-09-21.