GBM Models (1y / 3y)

Gradient boosting models predicting fixed-horizon stock returns using 464 engineered features. Displayed on the dashboard as GBM 1y and GBM 3y.

Overview

The GBM models are the core fixed-horizon ranking models, using comprehensive feature engineering and 8+ quarters of historical data to predict 1-year and 3-year forward returns. They use LightGBM with cross-sectional normalization to learn relative stock attractiveness.

Key Features

• Strong Accuracy: Rank IC 0.59-0.61
• Rich Feature Set: 464 features from 21 base metrics
• Data Requirements: 8+ quarters of history
• Coverage: ~52% of stocks (those with sufficient history)

Available Variants

GBM 1y

Predicts 1-year forward returns - Rank IC: 0.59 - Decile Spread: 75% - Top decile average return: 63% - Bottom decile average return: -12%

GBM 3y

Predicts 3-year forward returns - Rank IC: 0.61 - Decile Spread: Not specified (but strong) - Longer horizon allows mean reversion patterns

Feature Engineering

Base Features (21)

• Fundamentals: Profitability, growth, leverage, liquidity
• Valuation: PE, PB, PS, EV/EBITDA, EV/Revenue
• Market: VIX, Treasury rates
• Price: Returns, volatility, volume trends

Engineered Features (~22x per base)

1. Lag Features ([1, 2, 4, 8] quarters): - Captures historical values - Enables momentum patterns - Examples: pe_ratio_lag1q, revenue_growth_lag4q

2. Change Features: - QoQ: Quarter-over-quarter deltas - YoY: Year-over-year comparisons - Example: profit_margins_yoy

3. Rolling Statistics ([4, 8, 12] quarter windows): - Mean: Trend levels - Std: Volatility/stability - Slope: Direction/acceleration - Example: roe_mean8q, debt_to_equity_std12q

4. Missingness Flags: - Binary indicators for missing data - Captures data quality signal

5. Categorical: - Sector encoding (11 sectors)

Total: 464 Features

21 base + (21 × 4 lags) + (21 × 2 changes) + (21 × 3 stats × 3 windows) + 21 flags + 11 sectors
= 21 + 84 + 42 + 189 + 21 + 11 = 368 + overheads ≈ 464

Training Process

Cross-Sectional Normalization

# Per-date standardization
for date in unique_dates:
    features[date] = (features[date] - mean[date]) / std[date]

Why: Makes model regime-agnostic, focuses on relative rankings

LightGBM Configuration

params = {
    'objective': 'regression',
    'metric': 'rmse',
    'num_leaves': 31,
    'learning_rate': 0.05,
    'feature_fraction': 0.9,
    'bagging_fraction': 0.8,
    'bagging_freq': 5
}

Time-Series Cross-Validation

• 5-fold expanding window
• No data leakage across folds
• Preserves temporal ordering

Performance

Metric	GBM 1y	GBM 3y
Rank IC	0.59	0.61
Decile Spread	75%	Strong
Features	464	464
Min Quarters	8	8

GBM Lite (retired)

A simplified 59-feature variant (GBM Lite) was previously available with only 2-quarter data requirements. It was removed from the dashboard for producing overoptimistic predictions. See archived docs.

When to Use

Best For

• Maximum accuracy: When you need the best possible rankings
• Long-term holds: Extra accuracy matters more
• Established companies: 8+ quarters available
• Quantitative strategies: Systematic portfolio construction

Consider Other Models

• Absolute valuation: Use DCF/RIM instead of GBM
• Peak return timing: Use GBM Opportunistic or AutoResearch
• Broader coverage: AutoResearch covers more stocks with fewer data requirements

Feature Importance

Top Predictive Features (Typical)

1. Price Momentum (15-20% importance)
2. returns_3m, returns_6m, returns_1y

3. Strongest short-term signal

4. Valuation Changes (12-18%)

5. pe_ratio_qoq, pb_ratio_yoy

6. Direction of cheapening/expensive

7. Profitability Trends (10-15%)

8. profit_margins_slope4q, roe_mean8q

9. Quality improvement/deterioration

10. Growth Acceleration (10-12%)

11. revenue_growth_qoq, earnings_growth_slope

12. Second derivative matters

13. Volatility (8-10%)

14. volatility, roe_std8q
15. Risk-adjusted returns

Implementation

from invest.scripts.run_gbm_predictions import run_predictions

# Run GBM Full 1y
predictions = run_predictions(
    variant='standard',  # 'standard' = full model
    horizon='1y',
    db_path='data/stock_data.db'
)

# Get top quintile
top_20pct = predictions[predictions['percentile'] >= 80]

Theoretical Foundation

Why Historical Depth Matters

Mean Reversion Patterns: - High ROE tends to fade (rolling avg captures this) - Low margins tend to improve (slope detects acceleration) - Extremes revert to sector norms (std flags outliers)

Momentum Persistence: - 6-12 month price momentum predicts next 3-12 months - Fundamental momentum (growth acceleration) also persists - Lags capture these patterns

Volatility Regimes: - High volatility stocks underperform (risk penalty) - Volatility of fundamentals signals instability - Rolling std quantifies this

Cross-Sectional Learning

GBM learns relative attractiveness: - Ranking objective, not absolute returns - Per-date normalization removes market timing - Focus on stock selection within universe

Academic References

Gradient Boosting

• Friedman, J. H. (2001). "Greedy Function Approximation: A Gradient Boosting Machine". Annals of Statistics.
• Chen, T., & Guestrin, C. (2016). "XGBoost: A Scalable Tree Boosting System". KDD.
• Ke, G., et al. (2017). "LightGBM: A Highly Efficient Gradient Boosting Decision Tree". NIPS.

Factor Models & ML

• Gu, S., Kelly, B., & Xiu, D. (2020). "Empirical Asset Pricing via Machine Learning". Review of Financial Studies.
• Moritz, B., & Zimmermann, T. (2016). "Tree-Based Conditional Portfolio Sorts". Working Paper.
• Kozak, S., Nagel, S., & Santosh, S. (2020). "Shrinking the Cross-Section". Journal of Financial Economics.

Feature Engineering

• Jegadeesh, N., & Titman, S. (1993). "Returns to Buying Winners and Selling Losers". Journal of Finance.
• Fama, E., & French, K. (2015). "A Five-Factor Asset Pricing Model". Journal of Financial Economics.

Related Models

• AutoResearch: 5-model ensemble for peak 2-year return prediction
• GBM Opportunistic: Peak return prediction variant (1y and 3y)
• DCF: Absolute valuation alternative
• RIM: Residual income valuation for financials