An educational, end-to-end research platform for predicting five-session residual equity returns and constructing sector- and beta-neutral long-short portfolios. It is not investment advice.
point-in-time eligibility → leakage-safe features → nested OOS forecasts
→ neutral portfolio construction → execution and cost diagnostics
- Reproducible public-data and point-in-time-proxy pipelines
- Chronological, embargoed nested walk-forward model selection
- Market-neutral weekly portfolios with transaction-cost and capacity analysis
- Explicit research decisions: fragile signals are rejected rather than promoted
The completed public point-in-time membership-proxy study found positive baseline forecasts but insufficient robustness to claim alpha.
| OOS metric | Baseline entry | Two-session delay |
|---|---|---|
| Mean rank IC | 1.84% | 1.11% |
| Net Sharpe | 0.57 | 0.24 |
| Maximum drawdown | -7.02% | -8.93% |
| Mean weekly turnover | 76.98% | 77.63% |
At a $1m notional, baseline Sharpe falls from 0.34 at a 5 bps half-spread assumption to -0.02 at 10 bps and -0.73 at 20 bps. The study is therefore recorded as rejected, not presented as a tradable strategy.
python -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
python -m pytest -qpython -m scripts.data.download_pit_data --start 2016-01-01 --end 2025-12-31
python -m scripts.data.download_factors --output data/raw/fama_french_5_daily.parquet
python -m scripts.research.run_research \
--panel data/raw/sp500_pit_prices.parquet \
--benchmark data/raw/spy_benchmark.parquet \
--universe-history data/raw/sp500_pit_universe.parquet \
--factors data/raw/fama_french_5_daily.parquet \
--output data/processed/sp500_pit_proxy \
--model ridge --nested-validation --ridge-alphas 1,10,100 --test-days 252 \
--cost-model liquidity --cost-sensitivity-bps 5,10,20 \
--notional-sensitivity 1000000,5000000,10000000
python -m scripts.reporting.generate_report_assetsThe free pitindex/Yahoo workflow is a point-in-time membership proxy, not a
CRSP-quality data set: delisting returns, ticker history, corporate actions,
and price availability remain limitations. The generated run manifest records
parameters, input fingerprints, and environment information.
| Folder | Purpose |
|---|---|
src/ |
Research, modeling, portfolio, and backtest components |
scripts/data/ |
Public data, factor, and PIT-proxy download commands |
scripts/research/ |
Full-model and signal-library experiment commands |
scripts/reporting/ |
Figure and dashboard generation |
reports/ |
Dashboard, final report, data provenance, and decision log |
docs/ |
Research plan |
tests/ |
Leakage, alignment, neutrality, and cost controls |
For the detailed methodology, alternative workflows (SEC fundamentals and signal-library studies), and assumptions, start with the final report.