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

Episode Summary

Executive Summary: David Wright explains Pictet AM’s AI-enhanced quant process, emphasizing that it uses interpretable machine learning—primarily decision trees and gradient boosting on structured tabular data—rather than generative AI. The firm uses hundreds of signals across market, accounting, sell-side, and alternative data to forecast 20-day relative returns, then converts those forecasts into long-only ETFs aiming to match index risk while modestly outperforming by 1–2% annually.

Main Topics: AI vs. generative AI in quant investing (Priority: 5/5): Wright distinguishes traditional AI/machine learning from generative AI, arguing that GenAI is useful for support tasks but poorly suited for return forecasting because it can hallucinate and is hard to test, interpret, and keep free from look-ahead bias. Pictet’s machine-learning investment process (Priority: 5/5): The strategy uses thousands of decision trees trained via gradient boosting on structured data to forecast the next 20 days of relative stock performance across the investment universe. Data inputs and feature engineering (Priority: 5/5): The process combines hundreds of signals from prices, accounting data, analyst forecasts, positioning, calendar effects, and company-specific qualitative variables, with an emphasis on relationships and interactions among features. Where GenAI is actually used (Priority: 4/5): GenAI is being adopted for internal productivity—coding support, text interpretation, meeting summaries, and client materials—but not as the core alpha-generation engine. Portfolio construction and ETF objectives (Priority: 5/5): Forecasts are translated into long-only, risk-controlled portfolios that target benchmark-like beta and volatility while trying to deliver 1–2% annual outperformance in PQNT and PQUS. Interpretability and the ‘black box’ misconception (Priority: 4/5): Wright argues that quant and AI models can be transparent and explainable, with feature importance and tree structure helping identify what drives portfolio positions and performance.

Key Arguments: Generative AI is not well suited to forecasting returns because it is designed to generate outputs, can hallucinate, and is difficult to backtest without look-ahead bias. Decision-tree-based machine learning is preferred because it works well on tabular data, is testable, and produces interpretable forecasts. Quant investing has long used AI-like methods; the major recent change is wider adoption and cheaper/faster compute, not a fundamental shift in the core investment approach. Structured data and traditional financial signals remain central; text-based sentiment is being explored but has not yet clearly added enough value to displace existing signals. Machine learning can capture non-linear interactions among signals that traditional factor models often miss, allowing many more features to be included without the same rapid diminishing returns. Portfolio construction matters as much as signal generation because long-only ETF constraints prevent the model from fully expressing negative views, so the benchmark and risk controls must shape final positions. GenAI is already useful for administrative efficiency within the firm, but its contribution is mostly operational rather than alpha-generating. The firm views its products as passive replacements with an active edge: index-like risk plus modest excess return, rather than as market-neutral or alternative-return products.

Data Points: Assets under management in quant group: Over $30 billion - Pictet AM’s quantitative business, as described in the introduction. Number of ETF launches mentioned: 2 - PQNT (international equity) and PQUS (U.S. equity). Forecast horizon: Next 20 days - The model forecasts each stock’s relative performance over a 20-day period. Historical training window: 15 years - Feature data and outputs are assembled for every company on every trading day over 15 years. Number of signals/features: Hundreds / about 400 - Wright says the model uses roughly 400 input features drawn from multiple data sources. Signal mix: Roughly one-fourth each - Approximate breakdown among sell-side, market-based, accounting, and other data categories. Portfolio-model correlation: About 0.4 to 0.5 - Relationship between model views and active positions after benchmark and long-only constraints. Target excess return: 1% to 2% annually - Stated performance objective above the benchmark for both ETFs. Benchmark beta target: Beta of 1 - The ETFs aim to keep index-like risk characteristics and market exposure. Cloud spend trend: Reducing / lower than expected - Wright says model-training costs have fallen because training has become more efficient. ETF benchmarks: MSCI EAFE / S&P 500 - PQNT is benchmarked to MSCI EAFE (international developed markets ex-U.S./Canada) and PQUS to the S&P 500. Research runway before launches: 6 years - The U.S. ETF products are based on six years of research in the field.

Pivotal Quotes: "The last thing you want from any investor is information that is incorrect or that is being made up." — David Wright: Explaining why generative AI is not appropriate for return forecasting in quant investing. "The biggest one is the lack of understanding about what we're investing in. It's this idea of a black box." — David Wright: On the common misconception that quant models are opaque and unexplainable. "I think that efficiency is likely to mean that a lot of the expectations on the spend that are out there at the moment are very, very toppy." — David Wright: On whether the huge expected spending on AI infrastructure will translate into proportional returns.

Implications: The conversation suggests AI in investing will stay split: GenAI for productivity, interpretable ML for alpha. For investors, the key is judging whether a quant product truly adds signal and risk control, not just AI branding.