P2SAMAPA/p2-etf-rough-path-forecaster-results

P2 ETF Rough Path Forecaster Results

This dataset contains the output from the ROUGH-PATH-FORECASTER engine.

Engine Description

Uses signature kernel methods and Log-ODE for ETF return forecasting.

• —Signature Kernel: Neumann series expansion with dynamic truncation
• —Log-ODE: Neural controlled differential equations on log-signature space
• —Ensemble: Weighted combination of depths 2, 3, and 4

Universes

Fixed Income / Commodities

• —Benchmark: AGG
• —Tickers (7): TLT, LQD, HYG, VNQ, GLD, SLV, VCIT

Equity

• —Benchmark: SPY
• —Tickers (14): QQQ, XLK, XLF, XLE, XLV, XLI, XLY, XLP, XLU, XLRE, XLB, GDX, XME, IWM

Training Modes

Fixed Dataset

• —Period: 2008 → 2026 YTD
• —Split: 80% train, 10% validation, 10% test
• —Single model trained on all available data

Shrinking Windows (17 windows)

• —Start years: 2008 through 2024
• —End year: 2026 YTD (all windows)
• —Each window: independent model
• —Consensus scoring across windows

Consensus Weights

• —60% Annualized Return
• —20% Sharpe Ratio
• —20% (-)Max Drawdown

Output Structure

fi/ ├── fixed/ │ ├── model.pkl # Trained model │ ├── predictions.parquet # Test set predictions │ ├── actuals.parquet # Test set actual returns │ └── metrics.json # Performance metrics └── shrinking/ ├── modelwindow*.pkl # Per-window models ├── windowresults.parquet # Window metadata ├── consensus.parquet # Consensus pick ├── windowpicks.parquet # Per-window picks └── window_metrics.parquet # Per-window performance

equity/ └── (same structure as fi/)

metadata.json

Performance Metrics

Last Updated

2026-06-23T03:29:25.425342

License

MIT