ML Stock Selection Strategy

A per-sector ML ensemble that ranks the S&P 500 quarterly, wrapped in an MLOps pipeline: tracked training, an independently-verified backtest, containerized deployment, CI, and a live prediction dashboard.

Results

Backtested 2017-06-02 to 2026-08-05 (36 quarters, including the COVID crash and the 2022 bear market), selecting the top-3 performers per sector bucket:

	Annualized return	Sharpe	Max drawdown
Strategy	23.5%	0.93	-40.6%
SPY	15.2%	0.82	-33.7%
QQQ	20.0%	0.86	-35.1%

Strategy

The model is based on FinRL-Trading's own ml_bucket_selection.py: for each of 4 GICS-style sector buckets (growth_tech, cyclical, real_assets, defensive), 6 regressors (Random Forest, LightGBM, HistGradientBoosting, Extra Trees, Ridge, Stacking) compete on validation MSE to predict next-quarter return from 52 fundamental factors plus momentum, using point-in-time S&P 500 membership.

Architecture

flowchart LR
    subgraph vendor["vendor/FinRL-Trading (submodule)"]
        CSV[("fundamental_data_full.csv\n22,909 records / 715 tickers")]
        MODEL["ml_bucket_selection.py\n6 models x 4 sector buckets"]
        BT["BacktestEngine"]
    end

    CSV --> DB[("data/finrl_trading.db\nSQLite")]
    DB --> MODEL
    MODEL -- "predictions, feature importance,\nper-bucket val MSE" --> MLFLOW["MLflow tracking"]
    MODEL --> STRAT["MLBucketStrategy\n(BaseStrategy adapter)"]
    STRAT --> BT
    BT --> CMP["compare_vs_benchmarks.py\nvs. SPY / QQQ / equal-weight"]
    CMP --> JSON[("data/predictions/*.json")]
    JSON --> ST["Streamlit dashboard"]
    JSON --> NEXT["Next.js dashboard"]

    CRON["GitHub Actions\ndaily-selection.yml"] -.orchestrates.-> DB
    CRON -.-> MODEL

Loading

Repo structure

stock-selection/
├── vendor/FinRL-Trading/       # git submodule, upstream untouched
│   ├── src/strategies/ml_bucket_selection.py   # the model
│   ├── src/backtest/backtest_engine.py         # the backtest engine
│   └── data/                                   # bundled historical dataset
├── app/
│   ├── data/build_fundamentals_db.py           # CSV -> local SQLite
│   ├── strategy/ml_bucket_strategy.py          # BaseStrategy adapter
│   ├── training/run_selection.py               # runs the model, logs to MLflow
│   └── reporting/                              # portfolio math, frontend JSON export
├── backtests/
│   ├── run_backtest.py                         # strategy vs. SPY/QQQ
│   └── compare_vs_benchmarks.py                # + equal-weight control
├── streamlit_app/app.py                        # frontend, minimal tier
├── frontend/                                   # frontend, Next.js + Recharts
├── data/predictions/{latest,history}.json      # committed pipeline output, both frontends read this
├── Dockerfile / docker-compose.yml             # training image; MLflow + Postgres tracking backend
├── .github/workflows/{ci,daily-selection}.yml  # lint/test/backtest-smoke; scheduled retrain-and-rank
└── tests/

In progress: app/drift/ (monitoring, demoted to stretch), a FastAPI serving layer, notebooks/.

Getting started

git clone --recurse-submodules https://github.com/aluaz721/stock-selection.git
cd stock-selection
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# Build the local fundamentals DB from the bundled dataset
python -m app.data.build_fundamentals_db

# Run the model for one quarter and log it to MLflow
python -m app.training.run_selection --db data/finrl_trading.db --val-cutoff 2025-09-30

# Backtest across multiple quarters vs. SPY/QQQ/equal-weight
python -m backtests.compare_vs_benchmarks \
  --db data/finrl_trading.db \
  --infer-dates 2024-03-31 2024-06-30 2024-09-30 2024-12-31 \
  --top-n-per-bucket 3

Run the tests with pytest tests/, lint with ruff check ..

Frontends

Both read the same data/predictions/{latest,history}.json — regenerate with python -m app.reporting.build_frontend_data after a new backtest run.

# Streamlit
streamlit run streamlit_app/app.py

# Next.js
cd frontend && npm install && npm run dev

Tech stack

Python, scikit-learn, LightGBM, pandas, MLflow, bt (backtesting), SQLite, Streamlit, Next.js, React, TypeScript, Recharts, Tailwind CSS, Docker, GitHub Actions, pytest, ruff.

Limitations

• Max drawdown is worse than both benchmarks, consistently, including independently across two separate multi-year sub-periods.
• This project builds off of a published strategy. See https://github.com/AI4Finance-Foundation/FinRL-Trading

Full decision log, including things tried and reverted, in PROJECT_PLAN.md.

License

This repo's own code has no license file yet. vendor/FinRL-Trading/ is a separate upstream project (Apache 2.0), vendored unmodified.