Five risk components, each weighted not by a fixed rule but by a calibration loop that nudges weight toward the components whose realized outcomes beat the model's estimate. Move the market inputs and watch the score — and the weights — recompute.
Components scale raw inputs into 0–1 scores (VIX/40, OAS/300). Each learned weight = base + LR × base × (calibration signal − mean signal). The score is the learned-weight average of component scores.
Learning loop: each learned weight nudges toward components whose realized outcome exceeded the model estimate, scaled by the learning rate. Set the learning rate to zero to freeze learning at the base weights.
Composite = 0.35·reserve gap + 0.30·social inflation + 0.20·concentration + 0.15·liquidity · CRITICAL ≥ 0.60 · WATCH ≥ 0.40
60 seeded paths over 8 horizons · band = P05–P95 · line = median · paths are calibrated to a prior baseline estimate series, not to the live score above; shown to demonstrate the ensemble method pending recalibration
Efficiency = expected score impact per $100K of cost