The Alpha No Human Can Find | David Wright on Machine Learning's Hidden Edge

Key Concepts

• AI/ML for Idiosyncratic Alpha: Picet Asset Management utilizes Machine Learning (ML) to forecast stock-specific, idiosyncratic returns – the portion of return not explained by broad market factors – aiming for over 95% alpha generation.
• Decision Tree Approach & Interpretability: The core methodology employs interpretable decision tree models, allowing for tracing the reasoning behind forecasts, unlike “black box” neural networks.
• Feature Engineering & Data Prioritization: Approximately 400 features are used, prioritizing traditional data sources (fundamentals, pricing) due to historical depth, while remaining open to incorporating less-rationally explained features.
• Global Model Transferability: Extensive testing demonstrates the surprising stability and transferability of the model across different geographic regions and market types.
• Human-Machine Collaboration: The strategy emphasizes augmenting human expertise with ML, with humans defining relevant features and ML optimizing their weighting and interaction.

Leveraging AI for Stock Selection & Portfolio Construction

Picet Asset Management’s approach centers on using Machine Learning (ML) – a subset of Artificial Intelligence (AI) – to forecast the relative attractiveness of stocks over a one-month horizon. The focus isn’t on replacing human analysts, but on augmenting their abilities; humans are better at defining the universe of relevant features, while ML excels at determining the optimal weighting and interaction of those features. This is particularly effective for shorter investment horizons, where ML demonstrates a significant advantage over traditional factor models.

Model Methodology & Data Sources

The firm utilizes a decision tree-based ML approach. The model is trained on 15 years of historical data, updated every three months, and rigorously validated using out-of-sample testing and multiple training/validation splits to prevent overfitting. Approximately 400 features are employed, categorized as price-based, sell-side forecast-based, accounting-based, and investor positioning/calendar/qualitative data. While a purely data-driven approach is debated, the firm generally favors features with a strong theoretical or behavioral underpinning. A key differentiator is the interpretability of the decision tree models, allowing for tracing the reasoning behind each stock’s forecast.

Idiosyncratic Return & Portfolio Construction

The strategy focuses on generating stock-specific alpha by forecasting a stock’s idiosyncratic return – the portion of its return not explained by market, sector, industry, or country beta. This is achieved by statistically cleaning historical returns of these common beta factors. Portfolio construction employs “guard rails” to avoid significant deviation from benchmark weights, mirroring enhanced indexing, while aiming for over 95% of returns from stock-specific alpha. The strategy is currently deployed against the MSCI World benchmark in Europe and an EY benchmark in the US, with plans for expansion. A minimum universe size of 400-500 names is required for effective implementation, making a UK-specific strategy currently impractical.

Model Transferability & Rebalancing

Extensive testing revealed the remarkable transferability of the model across different geographic regions and market types. A globally trained model with a global feature set consistently outperformed models trained on specific regional data, suggesting relationships between features are remarkably stable. Portfolio rebalancing is currently done weekly, balancing the need to capture evolving model views with the minimization of transaction costs. The model updates daily, but weekly rebalancing is deemed optimal given the current strategy size and the high correlation of daily forecasts. A human makes the final rebalancing decision based on data analysis.

Complex Feature Interactions & Continuous Refinement

A surprising finding was the complex interplay of features – often six or more – working in combination to drive predictive power. Some relationships, like the interaction between calendar information and analyst forecasts, were intuitive, while many others lacked obvious explanations. The model also unexpectedly revealed a degree of “quality” exposure even when not explicitly targeted. The strategy deliberately avoids traditional factor construction, instead focusing on variations within factors (e.g., volatility in return on assets). The speaker emphasized the importance of a holistic approach to investment strategy, advocating for continuous refinement across all stages – feature generation, training, and portfolio construction.

Conclusion

Picet Asset Management’s AI-driven investment strategy demonstrates the potential of ML to generate stock-specific alpha by forecasting idiosyncratic returns. The success of the approach hinges on a combination of interpretable modeling techniques, a robust feature engineering process, and a commitment to human-machine collaboration. The surprising transferability of the model across markets and the complex interplay of features highlight the power of data-driven discovery, while the emphasis on continuous refinement underscores the dynamic nature of successful investment strategies. The advice to “don’t overtrade” reinforces the importance of aligning trading frequency with forecast horizons and minimizing transaction costs.