Decision Trees and Feature Splits Explained

In this episode of Excess Returns, we sit down with David Wright, Head of Quantitative Investing at Pictet Asset Management, for a deep and practical conversation about how artificial intelligence and machine learning are actually being used in real-world investment strategies. Rather than focusing on hype or black-box promises, David walks through how systematic investors combine human judgment, economic intuition, and machine learning models to forecast stock returns, construct portfolios, and manage risk. The discussion covers what AI can and cannot do in investing today, how machine learning differs from traditional factor models and large language models like ChatGPT, and why interpretability and robustness still matter. This episode is a must-watch for investors interested in quantitative investing, AI-driven ETFs, and the future of systematic portfolio construction.

Main topics covered:

• What artificial intelligence and machine learning really mean in an investing context

• How machine learning models are trained to forecast relative stock returns

• The role of features, signals, and decision trees in quantitative investing

• Key differences between machine learning models and large language models like ChatGPT

• Why interpretability and stability matter more than hype in AI investing

• How human judgment and machine learning complement each other in portfolio management

• Data selection, feature engineering, and the trade-offs between traditional and alternative data

• Overfitting, data mining concerns, and how professional investors build guardrails

• Time horizons, rebalancing frequency, and transaction cost considerations

• How AI-driven strategies are implemented in diversified portfolios and ETFs

• The future of AI in investing and what it means for investors

Timestamps:
00:00 Introduction and overview of AI and machine learning in investing
03:00 Defining artificial intelligence vs machine learning in finance
05:00 How machine learning models are trained using financial data
07:00 Machine learning vs ChatGPT and large language models for stock selection
09:45 Decision trees and how machine learning makes forecasts
12:00 Choosing data inputs: traditional data vs alternative data
14:40 The role of economic intuition and explainability in quant models
18:00 Time horizons and why machine learning works better at shorter horizons
22:00 Can machine learning improve traditional factor investing
24:00 Data mining, overfitting, and model robustness
26:00 What humans do better than AI and where machines excel
30:00 Feature importance, conditioning effects, and model structure
32:00 Model retraining, stability, and long-term persistence
36:00 The future of automation and human oversight in investing
40:00 Why ChatGPT-style models struggle with portfolio construction
45:00 Portfolio construction, diversification, and ETF implementation
51:00 Rebalancing, transaction costs, and practical execution
56:00 Surprising insights from machine learning models
59:00 Closing lessons on investing and avoiding overtrading