Financial-ML

A machine learning system for equity selection that predicts which S&P 500 stocks will outperform the market benchmark. The project combines market data with company fundamentals to construct long/short portfolios, using time series cross-validation to prevent lookahead bias and comprehensive backtesting to validate performance against SPY. The goal: identify stocks with positive alpha while maintaining robust risk management through systematic validation at both model and portfolio levels

Key Results

Metric	Value
Sharpe Ratio	0.93
Annual Return	20.2%
Max Drawdown	-22.9%
Alpha vs Random	1.72%
Win Rate	69.8%

• 100% Long-Only Strategy: Top 10% stocks, equal-weighted, monthly rebalancing
• Transaction costs included: 10 bps per trade, ~0.5% annual drag from 42% turnover
• Regime awareness: VIX-based features improve downside protection during volatile periods
• Statistically significant: Sharpe ratio outperformance (p < 0.001, Bonferroni-adjusted)

Pipeline Architecture

The pipeline of this project follows a standard MLOps workflow: data ingestion → feature engineering → model training with time series cross-validation → portfolio construction → backtesting.

flowchart TD
    A1[Sentiment Data] --> C[Feature Engineering]
    A2[Market Data] --> C
    A3[Fundamentals Data] --> C
    C --> D[Time Series CV & Training]
    D -.-> E[Model Evaluation]
    D --> F[Portfolio Construction]
    F -.-> G[Portfolio Diagnostics]
    F --> H[Performance vs SPY]

Loading

---

Table of Contents

• Features
• Project structure
• Installation
• Quick start
• Data Collection
• Modelling
• Evaluation
• Portfolio construction and backtesting
• Documentation

Features

• Fetch the current S&P 500 list, normalize tickers for the market data API, and persist the symbols table.
• Download monthly adjusted close prices for all S&P 500 tickers and the benchmark instrument, writing tidy CSVs for full or test universes.
• Retrieve selected fundamentals (assets, liabilities, equity, revenues, net income, EPS, shares outstanding) from public filings and save as a long-format CSV.
• Combine fundamentals and market information to produce ticker features: 13 features across market signals (momentum, volatility), fundamentals (value, quality, structure), and regime indicators (VIX percentile).
• Define an excess-return land evaluate baseline classifiers with expanding-window time series splits.
• Produce out-of-fold predictions per date and ticker for analysis and diagnostics.
• Use Random Forest predictions to construct 100% long-only portfolios (top 10% confidence) with monthly rebalancing and 3-month smoothing.
• Assess the model robustness and the qualities of the portfolios (e.g. Sharpe ratios, profits, risks, volatity vs SPY).

Project structure

The project is structured as follows (For a more detailed explanation please refer to docs/methodology.md ):

financial-ml/
├── src/financial_ml/ # Source code
│ ├── cli/ # Command-line interface components
│ ├── data/ # Data pipeline & feature engineering
│ ├── models/ # Training with time series CV
│ ├── evaluation/ # Model diagnostics
│ └── portfolio/ # Backtesting & construction
├── docs/ # Technical documentation
├── figures/ # Performance charts
└── requirements.txt

Key Modules

Data Pipeline (data/):

• Downloads market data (yfinance) and fundamentals (SEC EDGAR)
• Engineers 13 features across market, fundamental, and sentiment categories
• Handles missing data and forward-fills quarterly fundamentals

Model Training (models/):

• 3-fold expanding window time series CV
• Trains Random Forest with VIX-based regime features
• Saves trained models and predictions

Evaluation (evaluation/):

• Feature importance analysis
• Model diagnostics (AUC, calibration metrics)
• Cross-fold performance comparison

Portfolio Construction (portfolio/):

• Selects top 10% stocks by predicted probability
• Constructs 100% long-only portfolios with monthly rebalancing
• Backtests against SPY benchmark with comprehensive diagnostics

Output & Artifacts

After running the pipeline:

• Predictions: data/predictions/ - Out-of-fold predictions with probabilities
• Models: models/ - Trained model artifacts (.pkl files)
• Visualizations: figures/ - Portfolio performance charts, feature importance
• Logs: logs/ - Training and execution logs

For detailed directory structure, see docs/methodology.md.

Installation

Requirements: Python ≥ 3.10

Quick Install

# Clone repository
git clone https://github.com/pmatorras/financial-ml.git
cd financial-ml

# Create virtual environment (recommended)
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate

# Install package in editable mode
pip install -e .

Reproducible Install

For exact dependency versions used in development:

pip install -r Requirements.txt
pip install -e .

Note: The package will be installed as financial_ml and can be run via python -m financial_ml <command>.

Quick start

Complete ML Pipeline

Run the full pipeline from data collection to backtesting:

# 1. Collect market, fundamentals, and sentiment data
python -m financial_ml market
python -m financial_ml fundamentals
python -m financial_ml sentiment

#2. Train models
python -m financial_ml train

#3. Anayse training results
python -m financial_ml analyze

#4. Create portfolio and backtests
python -m financial_ml portfolio --model rf

Output: Models saved to models/, predictions to predictions/, and performance plots to figures/.

Faster alternative using the Makefile:

make data #collect all data
make train #train models only
make backtest #Analyse training results and do portfolio

Fast Development Mode

For quick iterations using a subset of ~50 stocks:

# Download test subset
python -m financial_ml market --test

# Train on subset (market features only, faster)
python -m financial_ml train --test --market-only

# Backtest
python -m financial_ml portfolio --model rf --test

Global Flags

Available for all commands:

--test

• Uses subset of ~50 stocks instead of full S&P 500
• Appends _test suffix to data files

--debug

• Enables verbose logging
• Designed for pipeline validation with minimal data (1-2 stocks)
• Saves artifacts to debug/ directories with _debug suffix

--do-sentiment

• Includes VIX-based sentiment features
• Improves regime awareness during volatile periods

--market-only

• Use only market features (excludes fundamentals)
• Faster for experimentation and baseline comparisons

Examples:

#Development mode (small subset)
python -m financial_ml train --test

#Debug mode (verbose logging + debug artifacts)
python -m financial_ml train --debug

#Combined for troubleshooting
python -m financial_ml train --test --debug

Tip: Use python -m financial_ml --help to see all available commands and flags.

Data Collection

Usage

# 1. Download S&P 500 constituent list and prices
python -m financial_ml market

# 2. Download company fundamentals from SEC EDGAR
python -m financial_ml fundamentals

#3. Download sentiment data
python -m financial_ml sentiment 

Command specific flags

• --newtable/-nt: Refresh S&P 500 constituent list from public source
• --newinfo/-ni: Redownload all historical price data (ignores cache)

Caching behaviour: By default, data files are only downloaded if they don't already exist. Use --newtable and --newinfo to force refresh.

Note: Market data collection and fundamentals collection are separate commands to enable independent execution. You can update constituents and prices without re-fetching fundamentals, or vice versa.

Data sources

Two types of data are currently considered: market and fundamental data:

1. Market Data (via yfinance)

   • Universe: S&P 500 constituents scraped from public reference
   • Prices: Monthly adjusted close prices for all symbols + SPY benchmark
   • Normalization: Ticker symbols standardized for API compatibility

2. Fundamental Data (via SEC EDGAR API)

   • Source: Company 10-K/10-Q filings
   • Tags: Selected US-GAAP financial metrics (revenues, assets, equity, etc.)
   • Processing: Point-in-time series, de-duplicated by metric/unit/date
   • Variants: Multiple tag variants to handle company-specific reporting differences

3. Sentiment data (via yfinance)

Note: Some fundamentals use different US-GAAP tag variants (e.g., Revenues, RevenueFromContractWithCustomerExcludingAssessedTax) to capture data across different reporting formats.

Target Variable

Binary classification: Predict whether a stock will outperform the S&P 500 benchmark.

Label definition: y = 1 if stock's 12-month forward return exceeds SPY's 12-month forward return, else y = 0.

This creates a relative momentum signal focused on identifying stocks that beat the market, suitable for long/short portfolio construction.

---

Modelling

Usage

# Train with fundamentals with sentiment (default)
python -m financial_ml train --do-sentiment

# With no sentiment data
python -m financial_ml train 

# Market data only (skip fundamentals)
python -m financial_ml train --market-only

Command specific tags

• --do-sentiment - Include VIX data (default).
• --only-market - Train using only market data (excludes fundamental ratios)
• --save - Saves model results.

Features

The model uses 13 features spanning market and fundamental factors:

Market Signals (4):

• Momentum: r12, mom121 (12-month trend)
• Volatility: vol3, vol12 (risk measures)
• Price: ClosePrice (behavioral signal)

Fundamental Ratios (9):

• Value: BookToMarket
• Quality: ROE, ROA, NetMargin
• Structure: Leverage, AssetGrowth, NetShareIssuance

Sentiment (1):

• Regime: VIXpercentile (12-month rolling percentile for market stress detection)

Note: ClosePrice and LogMktCap were tested but excluded—their information is already captured in returns and fundamental metrics.

Full feature definitions and formulas in docs/methodology.md.

Feature Importance:

Top features driving model predictions:

Feature	Mean Importance	Interpretation
vol12	25.8%	12-month volatility (quality/risk proxy)
BookToMarket	23.6%	Value factor
ROA	13.3%	Return on assets (profitability)
NetMargin	8.6%	Operating efficiency
VIXpercentile	0.23%	Regime indicator (critical in high-volatility periods)

Note: VIXpercentile has low average importance but provides critical protection during market stress. See docs/results.md for a more detailed analysis of feature roles across market regimes.

ML Models

Three binary classifiers predict monthly stock outperformance vs. SPY benchmark:

• Logistic Regression (L1) - Lasso regularization, C=0.5
• Logistic Regression (L2) - Ridge regularization, C=1.0
• Random Forest - 50 trees, max_depth=3, max_features='log2' ← Production model
• Random Forest (Calibrated) - Isotonic calibration (portfolio performance: worse than random)
• XGBoost, LightGBM, Gradient Boosting - Boosting variants (overfit in weak-signal regimes)

Configuration: See models/definitions.py for complete pipeline specifications.

Preprocessing pipeline:

• Sanitize infinite values → Replace with NaN
• Impute missing values → Median strategy
• Scale features → StandardScaler (linear models only)
• Balance classes → class_weight='balanced'

Full model specifications: src/financial_ml/models/definitions.py

Model Selection

Multiple models were evaluated using 3-fold time series cross-validation:

Model	Test AUC	Portfolio Sharpe	Result
Logistic Regression L1	0.558	~0.65	Baseline
Logistic Regression L2	0.558	~0.65	Baseline
Random Forest	0.525	0.93	✅ Production
Random Forest (Calibrated)	0.526	0.79	Worse with VIX
XGBoost	0.524	0.80	Overfits
LightGBM	0.526	0.78	Overfits
Gradient Boosting	0.526	0.84	Overfits

Why Random Forest Wins:

• Random forest generally outperforms boosting in weak-signal regimes (AUC ≈ 0.525)
• Maintains wider probability spread for portfolio selection
• Naturally learns regime-dependent interactions with VIXpercentile

See docs/results.md and docs/experiments.md for detailed analysis.

Cross-Validation

3-fold expanding window time series CV to mimic production and prevent lookahead bias and data leakage:

Fold	Train Period	Test Period	Test AUC
1	2010-2016 (6 yrs)	2016-2018 (2 yrs)	0.18
2	2010-2018 (8 yrs)	2018-2021 (3 yrs)	0.517
3	2010-2021 (11 yrs)	2021-2025 (4 yrs)	0.541

Average Test AUC: 0.525

---

Evaluation

One can analyse trained model feature importance and coefficients without retraining, simply by retrieving the models training information.

Usage

python -m financial_ml analyze --model all #or any model in definitions.py

Output

• Feature importance plots - Random Forest feature rankings saved to figures/
• Coefficient plots - Logistic Regression coefficients saved to figures/
• Model correlation matrix - Heatmap showing prediction agreement across all models

What's Analysed

• Random Forest - Feature importance (Gini impurity reduction)
• Logistic Regression - Coefficient magnitudes and signs
• Top features - Ranked by contribution to predictions

Implementation: See evaluation/feature_analysis.py for details

---

Portfolio Construction and backtesting

Construct 100% long portfolios from model predictions and evaluate performance against SPY benchmark.

Usage

python -m financial_ml portfolio --model rf # Or any other model in definitions.py

Strategy:

• Long top 10% of stocks by predicted probability
• Equal-weighted positions within each leg
• Rebalancing Monthly (last trading day)
• Smoothing 3-month rolling average to reduce noise and turnover

Performance Metrics

Metric	Description
Cumulative Return	Total portfolio return over backtest period
Sharpe Ratio	Risk-adjusted return (annualized)
Max Drawdown	Largest peak-to-trough decline
Turnover	Average monthly portfolio rebalancing
Win Rate	% of months outperforming SPY
Turnover	Average monthly portfolio churn
Beta to SPY	Market exposure and correlation

Diagnostics

• Sector drift - Portfolio sector allocation vs SPY over time
• Model correlation - Agreement between different model predictions
• Regime performance - Returns during calm vs volatile market periods

Output

• Performance charts - Cumulative returns, drawdown and sector drieft plots saved to figures/
• Backtest results - Metrics printed to console

Implementation: See portfolio/backtest.py

---

Documentation

This README provides a high-level overview. For detailed information, please refer to:

• docs/experiments.md - Complete ablation studies, rejected approaches, lessons learned
• docs/methodology.md - Technical implementation details, formulas, model specifications
• docs/decisions.md - Design rationale for key choices (features, rebalancing frequency, etc.)

---

Back to top