Factor Forge

A modern, reproducible factor zoo replication engine for cross-sectional equity research.

Factor Forge takes the academic "factor zoo" — momentum, value, quality, low-volatility, investment, profitability — and makes it runnable, honest, and ML-ready. It handles transaction costs, survivorship bias, look-ahead bias, factor decay, and walk-forward machine learning.

Mission: Let any researcher or student answer "Does this factor still work after costs?" with net-of-cost Sharpe, turnover analysis, and uncertainty-aware ensembles.

---

Why this matters

• Quantopian is gone and zipline / pyfolio are unmaintained.
• Most public factor code reports gross returns, ignores transaction costs, and suffers from look-ahead bias.
• Factor Forge reports net-of-cost Sharpe, turnover, capacity, and factor decay in a clean, extensible Python package.

---

What you can do

• Compute 15+ classic factors with a uniform cross-sectional API.
• Build decile long/short and long-only portfolios with monthly rebalancing.
• Apply a transaction-cost and slippage model to every trade.
• Measure information coefficient (IC), factor decay, turnover, and regime performance.
• Train a walk-forward LightGBM ensemble that combines factors without peeking.
• Add conformal prediction intervals for return forecasts.
• Generate publication-ready plots, interactive HTML research reports, and a Streamlit dashboard.

---

Quickstart

git clone https://github.com/satyamdas03/factor-forge.git
cd factor-forge
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -e ".[dev]"

Or install from PyPI:

pip install factor-forge-quant

Run a backtest on a single factor using cached or Polygon data:

export POLYGON_API_KEY=your_key_here
factor-forge --factor momentum_12_1 --start 2015-01-01 --end 2024-01-01

Generate an HTML research report:

ff-report --factor momentum_12_1 --start 2015-01-01 --end 2024-01-01 --output reports

Launch the interactive dashboard:

pip install "factor-forge-quant[dashboard]"
ff-dashboard

Or from Python:

from factor_forge import BacktestEngine, DataLoader
from factor_forge.factors.registry import FactorRegistry

loader = DataLoader(source="polygon", api_key="your_key")
prices = loader.load_eod(["AAPL", "MSFT", "GOOGL"], start="2015-01-01", end="2024-01-01")

registry = FactorRegistry()
registry.load_builtin()
factor = registry.get("momentum_12_1")

engine = BacktestEngine(factor=factor, prices=prices, transaction_cost_bps=10.0)
result = engine.run()
print(result.metrics)

---

Factor library

Category	Factors
Momentum	momentum_12_1, momentum_52_high, residual_momentum
Value	pe_ratio, pb_ratio
Quality	roe, gross_margin, accruals
Low volatility	idiosyncratic_volatility, beta, max_drawdown
Investment	asset_growth, capex_growth
Profitability	gross_profits_to_assets

Each factor is a pure function prices / fundamentals → cross-sectional score. Adding a new factor is one function + one registry entry.

---

Architecture

factor_forge/
├── data/          # Data sources, loader, cache, survivorship handling
├── factors/       # Factor definitions and registry
├── backtest/      # Portfolio, cost model, decile engine
├── analytics/     # Metrics, IC, decay, turnover, regimes, correlations
├── ml/            # Walk-forward features, LightGBM ensemble, conformal prediction
├── viz/           # Publication-ready plots
├── cli/           # Command-line tools
└── docs/          # MkDocs site + blog

---

Roadmap

Version	Goal	Status
v0.1.0	Data pipeline + 5 core factors + decile backtest	✅
v0.2.0	15+ factors + transaction costs + analytics	✅
v0.3.0	Factor decay, turnover, regime analysis	✅
v0.4.0	ML ensemble + conformal prediction	✅
v0.5.0	Dashboard + research report generation	✅

---

Documentation

Full documentation, API reference, and research blog are at:

https://satyamdas03.github.io/factor-forge/

---

Interview talking points

• "Most retail factor backtests report gross returns. Factor Forge reports net-of-cost Sharpe and turnover-adjusted returns."
• "I handle delisting and point-in-time fundamentals to avoid look-ahead bias."
• "I found momentum still works but turnover is high; low-vol is crowded and decayed post-2020."
• "I combined factors with a walk-forward LightGBM ensemble and added conformal prediction intervals for risk."

---

License

MIT © Satyam Das