trading-hub · calibration · d82 decision loop

d82 Decision-Feedback Loop

Every published THv2 decision (signal) has a predicted prob_up. As outcomes resolve (hit/miss), the Beta-Binomial conjugate posterior accumulates evidence about the realized hit probability per (tier × horizon) cohort. The cohort is calibrated when the predicted prob_up sits inside the 95% credible interval. When it doesn't, that cohort is surfacing predictably mis-calibrated probabilities — and Kelly sizing on that cohort should freeze until the gap closes.

Phase B-2 shipped: The “predicted” column now reports the tier-aware predicted hit rate Φ-derived from predicted_drift and predicted_horizon_sd, not the raw predicted_prob_up. This is an apples-to-apples comparison with the posterior. Hover the value to see which probe is in use (tier_aware_hit_rate for B-2 records; prob_up_fallback for legacy records seeded before B-2).

Loading d82 calibration state…
Total decisions
Resolved
Pending
Last cron run

Cohort-level posteriors

Pooled Beta(α, β) per (tier × horizon). α = 1 + Σ hits, β = 1 + Σ misses. 95% credible interval from regularized incomplete-beta inverse. Calibration verdict: predicted prob_up ∈ [lower, upper] ⇒ ✓; outside ⇒ ✗.

Tier Horizon n_obs μ̂ (posterior) 95% CI Predicted Δ Calibrated?
Loading…