Introduction

Fundamental analysis examines company economics: earnings, cash flow, valuation. Technical analysis examines price and volume patterns. In modern quant research, the dichotomy is false: machine learning naturally combines heterogeneous signals. This article explores hybrid frameworks incorporating both fundamentals and technicals.

The Case for Hybrid Approaches

Pure fundamental models sometimes miss short-term momentum. Pure technical models ignore valuation, creating crashes when mean-reversion finally occurs. Hybrid models combining both capture value creation (fundamentals) and market dynamics (technicals).

Feature Engineering: Fundamentals + Technicals

Fundamental Features

Earnings yield (earnings / price), book-to-market ratio, ROE (return on equity), debt-to-equity, dividend yield, earnings surprise, analyst estimate revisions, insider trading activity.

Technical Features

Price momentum (recent return), moving average deviations, volatility, RSI (relative strength index), MACD (moving average convergence divergence), volume patterns, support/resistance levels.

Combined Features

Ratios combining both: valuation (fundamental) relative to price momentum (technical), e.g., "cheap and rising," "expensive and rising," "cheap and falling." These interaction terms capture regime-dependent relationships.

Modeling Architectures

Ensemble Approach: Separate Branches

Train separate models on fundamentals and technicals, then ensemble. Fundamental model outputs probability stock outperforms fundamentally; technical model outputs probability based on technicals. Average predictions or use meta-learner to weight.

Advantage: separate models are interpretable. Disadvantage: doesn't capture interactions between fundamental and technical factors.

Deep Learning Fusion

Use neural networks with multiple input branches: fundamental branch processes fundamental features, technical branch processes technical time series. Branches converge to shared layers, then final prediction layer. Network learns relationships between fundamental and technical factors.

Gradient Boosting with Mixed Features

XGBoost and LightGBM handle mixed continuous features naturally. Create single feature matrix combining fundamentals and technicals. Tree-based models automatically identify splits that matter, discovering that "cheap + rising" is powerful signal while "cheap + falling" is weak signal.

Addressing Temporal Mismatch

Fundamental data (earnings, balance sheet) updates quarterly; technical data (price, volume) updates daily. When combining, must handle different frequencies carefully.

Options: (1) resample technicals to quarterly frequency (loses information), (2) forward-fill fundamentals (assumes no change until next earnings), (3) use separate models for quarterly rebalance (fundamentals) and daily trading (technicals).

Case Study: Value Plus Momentum

Simple hybrid: "Buy stocks that are cheap (low P/E ratio) and rising (positive momentum)." Separately, cheap and falling stocks underperform, and expensive rising stocks often crash. The combination matters.

In regression: next_return ~ value_score + momentum_score + (value_score * momentum_score). Interaction term captures non-additive effect: value and momentum together predict better than sum of parts.

Regime-Dependent Model Weights

Value factors perform well in some regimes (recession recoveries) and poorly in others (growth-stock dominated markets). Rather than single model, implement regime-switching: identify market regime, then use model weights appropriate for regime.

In low-rate environments, value is weak and growth/momentum strong. In high-rate environments, value is strong. Detect rate regime and adjust fundamental/technical weights accordingly.

Avoiding the Pitfalls

Data Leakage

Fundamental data is often released with delay (earnings 4-6 weeks after quarter end). When backtesting, use only fundamental data known at decision time, not future data.

Overfitting to Multiple Signals

Combining many signals (dozens of fundamentals, dozens of technicals) enables overfitting. Use feature selection: identify truly predictive signals, discard redundant ones. Cross-validate carefully.

Regime Shifts

Relationships between fundamentals and technicals aren't stable. During COVID, technicals dominated (fear-driven). During normal times, fundamentals matter more. Be prepared for model degradation and retrain regularly.

Implementation Strategies

Rolling Window Retraining

Retrain model monthly or quarterly on recent 3-5 years data. This adapts to changing relationships while maintaining sufficient training data.

Online Learning

Use models supporting incremental updates (online gradient descent, online tree learning). Each day, observe new data point, update model. More responsive to changing patterns.

Ensemble Across Time Periods

Train multiple models on different time periods, ensemble their predictions. Guards against single time period's patterns being anomalous.

Conclusion

The dichotomy between fundamental and technical analysis is outdated. Modern machine learning naturally combines heterogeneous signals. Hybrid models capturing both value creation (fundamentals) and price dynamics (technicals) outperform using either alone. Success requires careful feature engineering, proper handling of temporal mismatches, and continuous vigilance for overfitting and regime shifts. Best practitioners treat fundamental and technical signals as complementary inputs to machine learning systems rather than competing philosophies.