Model family

Gradient-boosted quantile regression (LightGBM). For each instrument, 5 independent regressors are trained — one per quantile level (q0.10, q0.25, q0.50, q0.75, q0.90) — each minimizing the pinball loss for its own quantile, jointly producing a full predicted price distribution rather than a single point forecast. Hyperparameters: n_estimators=200, max_depth=4, learning_rate=0.05, subsample=0.8, colsample_bytree=0.8.

Input features

Two families of engineered features, multiplicatively interacted:

1. Each instrument's own multifractal price dynamics — trace-moment parameters, structure-function exponents, and correlated/decorrelated structure-function gaps, computed on a trailing 512-trading-day rolling window and cross-sectionally z-scored against a peer group of comparable instruments (reused from this program's Paper 11 multifractal-predictability research).
2. Two forward-looking market regime signals: a credit-spread regime (HYG/LQD high-yield-to-investment-grade bond ratio, 200-day rolling z-score) and a VIX-term-structure regime (VIXM/VIXY medium-vs-short-term volatility-ETP ratio, 200-day rolling z-score) — both independently Granger-causality-validated against realized volatility before use, not included on correlation alone.

Model selection: four candidates compete, per instrument

Candidate	What it is
Climatology	A frozen day-of-year empirical quantile table — no machine learning at all, the pure calendar baseline
Credit-Regime	LightGBM on multifractal features × credit-spread regime
VIX-Regime	LightGBM on multifractal features × VIX-term-structure regime
Credit+VIX	LightGBM using both regime interactions together

The winner for each instrument is whichever candidate has the lowest mean absolute percentage error (MAPE) on a genuinely held-out period (2022 onward) never touched while choosing between candidates — climatology is not a baseline to beat, it is a real candidate, and it wins outright for 12 of the 22 instruments shown below.

Training: research vs. live

The backtested research behind this dashboard used strict walk-forward validation (each model trained only on data before the period it was tested on). The live forecasts shown here instead retrain a single final model on all available history through today, each time this page is rebuilt, and predict forward from today's freshest feature values — the one genuinely new mode of operation versus the backtested research.

Post-processing (2 of 22 instruments only)

For GLD and JPM specifically — the only two instruments, out of five independently designed correction techniques tested, where a real and repeatable forecast bias was found — an ongoing rolling bias correction (moment-matching then quantile-mapping, refit from a trailing window of already-resolved past predictions) is layered on top of the raw model output. The other 20 instruments use the raw model forecast, uncorrected, because correction was tested and found to make them worse.

Full methodology

Every equation, all five trading-strategy tests, all five post-processing designs, and the honest economic results (including why none of this is shown as a trading signal) are documented in the accompanying draft preprint. Code: notebooks/predictor_v1/.