$ cat methodology.md

Conformal prediction is a distribution-free framework for constructing prediction sets with finite-sample coverage guarantees. Given a calibration set {(X_i, Y_i)}_{i=1}^n and a new test point X_{n+1}, the conformal prediction set C(X_{n+1}) satisfies P(Y_{n+1} ∈ C(X_{n+1})) ≥ 1 - α under the exchangeability assumption. No parametric distributional assumptions are required. We implement three conformal methods: StandardCP (split conformal with quantile-based thresholds), MondrianCP (group-conditional coverage for regime-stratified data), and CQR (Conformalized Quantile Regression for heteroscedastic intervals).

StandardCP: Split conformal with nonconformity scores s_i = |Y_i - μ̂(X_i)| for regression or 1 - π̂_y(X_i) for classification. The prediction set includes all labels y where the score is below the (1-α)(1+1/n) quantile of calibration scores. MondrianCP: Group-conditional conformal that partitions the calibration set by regime labels (EXPANSION/CONTRACTION/TRANSITION) and computes per-group thresholds. Provides coverage guarantees within each group, not just marginally. CQR: Trains quantile regressors at levels α/2 and 1-α/2, then conformally calibrates the residuals to guarantee finite-sample coverage, producing intervals that adapt to local variance.

The Hamilton filter regresses Y_t on Y_{t-h} (typically h=24 months) to extract the cyclical component as the residual: ε_t = Y_t - β̂₀ - β̂₁Y_{t-h}. This avoids the spurious cycle artifacts and endpoint instability of the Hodrick-Prescott filter. The residuals preserve the spectral properties of the underlying cyclical dynamics without introducing phase shifts.

The Lomb-Scargle periodogram estimates spectral power P(ω) at angular frequency ω from unevenly sampled data. We apply it to Hamilton-filtered residuals to identify the dominant cycle frequency. The fitted sinusoid f(t) = A·sin(ωt + φ) parameterizes the cycle position into four quadrants: TROUGH (φ ∈ [π, 3π/2)), RECOVERY (φ ∈ [3π/2, 2π)), PEAK (φ ∈ [0, π/2)), CONTRACTION (φ ∈ [π/2, π)). R² of the sinusoidal fit quantifies cycle regularity.

Net S-D is a composite z-score: z(supply) - z(demand). Supply components (e.g., Fed total assets, M2, bank reserves) are individually z-scored over a trailing window and averaged. Demand components (e.g., Treasury general account, RRP facility balance) are similarly z-scored, optionally sign-inverted, and averaged. The composite is computed monthly from 1998–present. Year-over-year changes in Net S-D drive regime classification: sustained positive ΔNet S-D → EXPANSION, sustained negative → CONTRACTION, ambiguous → TRANSITION.

Regime labels are assigned based on year-over-year ΔNet S-D with hysteresis thresholds to prevent chattering. The classification uses a 3-state model: EXPANSION (ΔNet S-D > +σ for ≥3 consecutive months), CONTRACTION (ΔNet S-D < -σ for ≥3 consecutive months), TRANSITION (otherwise). MondrianCP leverages these regime labels for group-conditional conformal inference, providing coverage guarantees stratified by market state. Regime durations range from 4 to 36 months historically.

Temporal cross-validation with purged groups: fold boundaries are determined by time-sorted indices with a gap buffer equal to the forecast horizon. Within each fold, the model is trained on data before the purge boundary and evaluated on data after it. Coverage is measured as the fraction of test points whose true value falls within the conformal prediction set. We target 87.5% coverage (α=0.125) with an acceptance band of [85%, 90%]. Slots are classified as ON_TARGET (85-90%), OVERCOVERING (>90%), or UNDERCOVERING (<85%).

Macro regime context improves intraday alpha through three mechanisms: (1) Directional bias: BTC 12-month returns average +127% in EXPANSION vs -31% in CONTRACTION regimes since 2015, providing a statistical prior for trade direction. (2) Volatility regime: VIX z-score > 1.5 correlates with 2.3x average true range expansion, affecting position sizing and stop placement. (3) Correlation regime: Cross-asset correlations shift significantly between regimes (BTC-SPX correlation: +0.62 in contraction vs +0.21 in expansion), affecting portfolio hedging and pair trades.

FRED API ingestion covers 281 series across 6 categories: Liquidity (47 series), Inflation (31), Credit (28), Equity (42), Commodities (35), FX (19). Each series retains full history. Market data from Yahoo Finance covers 8 core assets at daily resolution. Derived features include z-scores (trailing 36-month window), year-over-year changes, Hamilton-filtered residuals, and Lomb-Scargle cycle parameters. Feature engineering produces ~40 derived features per series. Total feature space: ~11,240 features before selection.