Work In Progress

---

Factor Weaver

Master's thesis developing a reinforcement learning framework for portfolio management that integrates price, technical, fundamental, and behavioral (attention & sentiment) features across a dynamic equity universe of the top 50 S&P500 constituents. Uses an actor-critic architecture trained via PPO to output portfolio weights (no short-selling) maximizing risk-adjusted returns (Sharpe) under transaction costs, evaluated against benchmarks like equal-weight and minimum-variance.

Tech stack

Python, PyTorch, Gymnasium, PPO (hand-rolled), cross-stock transformer, Dirichlet policy head. Config via YAML, experiment tracking via MLflow, linting with ruff and pyright, tests with pytest.

Repository structure

├── config/             # YAML configs (data paths, env params, model hparams)
├── data/               # symlinks → /mnt/usb/factor-weaver/data/ (raw, interim, processed)
│   ├── raw/            #   original downloads/API output from vendors
│   ├── interim/        #   cleaned and normalized data in parquet files
│   └── processed/      #   final aligned output for RL training/testing
├── docs/               # thesis planning docs
│   └── figures/        #   PlantUML diagrams (data flow, architecture)
├── notebooks/          # exploratory quarto notebooks
├── src/factor_weaver/
│   ├── cli.py          #   argparse dispatcher (entry point)
│   ├── workflows/      #   orchestrators: build_dataset, train, evaluate
│   ├── data/           #   MODULE 1: data pipeline (parse, lseg, universe, edgar, prices, technicals, behavior, align, split)
│   ├── rl/             #   MODULE 2: env, model, ppo
│   ├── eval/           #   MODULE 3: backtest, benchmarks
│   └── math.py         #   financial math helpers (shared)
├── experiments/        # run outputs (logs, checkpoints, results)
├── tests/

Data pipeline

1. parse-fundamentals — financialdatadb xlsx → long parquet
2. parse-companies — financialdatadb company list → parquet
3. lseg-constituents — current S&P500 snapshot (LSEG chain RIC)
4. lseg-joiners-leavers — S&P500 membership changes since 1994 (LSEG)
5. lseg-mapping — RIC → ticker/name/PermID crosswalk for all LSEG RICs
6. lseg-market-cap — quarterly market cap via TR.CompanyMarketCap (fallback TR.F.MktCap)
7. universe — top-50 S&P500 by market cap per quarter
8. edgar-filing-dates — SEC filing dates (no look-ahead bias)
9. fetch-prices — LSEG OHLCV
10. compute-technicals — rolling indicators from prices
11. load-behavior — sibling-repo sentiment/attention parquet
12. align — EDGAR-anchored forward-fill + freshness + time-decay
13. split — train/test tensors by date window

Architecture and data-flow diagrams: docs/figures/.