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.

flowchart LR
  sources["sources<br/>LSEG · Yahoo · MarketPsych"] --> data["data pipeline<br/>fetch → build"]
  data --> panel[("panel + prices<br/>dynamic top-50 universe")]
  panel --> models["models<br/>RL policy + benchmarks"]
  models --> engine["shared daily<br/>backtest engine"]
  engine --> results["metrics · equity curves<br/>MLflow · report"]

Loading

Status

• Implemented: data pipeline through prices (LSEG universe/prices, Yahoo fallback + extra series), config, CLI and workflows (data, report), tests.
• Stubbed: technicals, panel, split; all models; shared backtest engine; train and eval.

Tech stack

• Implemented: Python, pandas, pyarrow, lseg-data, yfinance, scikit-learn, PyPortfolioOpt, MLflow, YAML configs, ruff + pyright, pytest.
• Planned: PyTorch, Gymnasium, hand-rolled PPO, cross-stock transformer, Dirichlet policy head.

Usage

uv sync                                    # create/update .venv from uv.lock
uv run factor-weaver data                  # run all steps (fetch then build)
uv run factor-weaver data fetch [steps]    # raw data acquisition only (cached)
uv run factor-weaver data build [steps]    # derived datasets only
uv run factor-weaver data --list           # step/output status
uv run factor-weaver data fetch --refresh  # refetch cached raw data
uv run factor-weaver train rl              # train the RL policy (stub)
uv run factor-weaver eval [models ...]     # backtest models via the shared engine (stub)
uv run factor-weaver report                # compare tracked eval runs (MLflow)

Run from the repository root so the relative config/ and data/ paths resolve. Fetch steps are existence-cached (whole outputs and per-RIC/per-symbol price caches); -v logs every cache hit/fetch.

Documentation

• docs/architecture.md — modules, data pipeline, runtime flow
• docs/data.md — sample, variables, sources
• docs/methodology.md — environment, model architecture, evaluation
• docs/models.md — benchmarks + shared backtest engine
• docs/literature.md — annotated references