📐 Alpha Calibration — Self-Improvement Loop

Self-Improvement Loop · Institutional Quant Diagnostics

Reads the trade journal weekly and recomputes optimal alpha-score weights using Information Coefficients, factor attribution OLS regression, Wilson confidence intervals, and Bayesian shrinkage with hard guardrails. Champion/challenger A/B with manual approval default. After 60+ trades, weight changes become statistically defensible.

⚖ Weight Evolution current vs proposed · shrinkage + guardrails applied —

CURRENT (in production)

PROPOSED (this calibration)

🎯 Per-Strategy Performance Sharpe · Sortino · expectancy · profit factor · t-stat · p-value · vs SPY

⚡ Information Coefficients corr(component_score_at_call, forward_return) at 1d / 7d / 30d / 90d

Component	IC 1d	n / sig	IC 7d	n / sig	IC 30d	n / sig	IC 90d	n / sig

🔬 Factor Attribution OLS return_30d ~ alpha_components

Factor	Coefficient (β)	Std Error	t-stat	p-value	Significance

📐 Methodology Reference how this calibrator works

PER-STRATEGY STATISTICS

sharpe_annualized = (mean_return / std) × √12 · sortino = (mean / downside_std) × √12
expectancy = win_rate × avg_win + loss_rate × avg_loss
profit_factor = sum(wins) / |sum(losses)|
wilson_ci_95 = (p + z²/2n ± z·√(p(1-p)/n + z²/4n²)) / (1 + z²/n) where z=1.96
t_stat = (mean - 0) / (std / √n) · p = two-tailed t-distribution

INFORMATION COEFFICIENTS

IC_h = corr(component_score_at_call, return_h) for h ∈ {1d, 7d, 30d, 90d}
IC standard error = √((1 - IC²) / (n-2)) · t = IC / SE

FACTOR ATTRIBUTION OLS

return_30d ~ β₀ + β₁·quality + β₂·growth + … + β₈·options_flow
β = (X'X)⁻¹ X'y · coefficients centered around component=50
data_implied_weight[i] = max(0, β[i]) / Σ max(0, β[j])

BAYESIAN WEIGHT UPDATE

proposed = (1 - λ) × current + λ × data_implied
λ = min(N / 200, 0.40) · capped shrinkage
guardrail: |Δweight| ≤ 0.03 per cycle · floor 0.04 · ceiling 0.22
renormalize: Σ weights = 1.00

DEPLOYMENT GATES

auto-apply requires ALL of:   1. auto_apply_calibrations flag = true (manual approval default)
  2. n_obs ≥ 60 evaluated 30d outcomes
  3. At least one factor with attribution p < 0.05
Otherwise: weights logged as "proposed" only, alpha-score uses current.

📊 Calibration History rolling 52-week audit log

Date	Trades	Win 30d	Avg Return	R²	Decision