Why Generative AI Still Can’t Trade | David Wright on How Quant Alpha Actually Is Done With Machine Learning, Decision Trees, and Gradient Boosting

Despite the hype around generative AI, David Wright of Pictet Asset Management argues that it remains fundamentally unsuited for alpha generation in quantitative investing. Drawing on over $30 billion in managed assets and six years of research, Wright explains that his team relies on interpretable, decision-tree-based machine learning—specifically gradient boosting—rather than generative AI, which he calls 'a system that makes things up.' The core issue, he says, is that generative AI introduces look-ahead bias during backtesting due to its reliance on vast, uncontrolled text data, making it unreliable for forecasting returns. Instead, Pictet's strategy uses 400+ structured, tabular signals—ranging from market trends and accounting changes to analyst sentiment and calendar effects—trained on 15 years of daily data to forecast 20-day relative performance. The model’s strength lies in uncovering complex, non-linear relationships between signals, such as how analyst recommendations lose predictive power as earnings reports approach. Crucially, the process is transparent: the 'crystal box' approach allows full interpretability of why positions are taken. While generative AI is used for internal efficiency—like drafting client decks—Wright warns that expectations of massive cloud spending and AI-driven returns are overblown, noting that his firm’s model training has become more efficient, not more expensive.