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

Podcast: Excess Returns
Date: December 17, 2025
Guest: David Wright, Head of Quantitative Investing at Pictet Asset Management
Hosts: Jack Forehand, Justin Carbonneau, Matt Zeigler

Episode Overview

This episode explores the integration of artificial intelligence (AI) and machine learning (ML) into quantitative investing, guided by the practical experience of David Wright. As head of quant investing at Pictet Asset Management, Wright discusses how ML models, particularly decision tree-based methods, are transforming stock selection, portfolio construction, and the overall investment process. The conversation covers the strengths and limitations of machine-driven strategies, the complementary roles of human judgment, and how technological advances might shape the future of systematic investing.

Key Discussion Points & Insights

1. Defining AI & Machine Learning in Finance

• AI vs. Machine Learning:
  • AI is any machine-driven activity mimicking human tasks, but in finance, most "AI" refers to machine learning—specifically, algorithms trained on data to autonomously learn relationships ([03:54]).
  • Quote: “AI, artificial intelligence, I simply think it means a machine doing some kind of function or role that a human would generally do… In practice though, generally what we’re talking about...is machine learning.” —David ([03:54])
• Training ML Models:
  • The key is selecting input “features” (signals about stocks/companies) and an output to train against (typically forward returns).
  • Rather than assigning weights to features by hand, models are fed historical data, which they use to “learn” the optimal way to forecast returns ([05:05]).

2. Machine Learning Models vs. Large Language Models (LLMs)

• LLMs (e.g., ChatGPT) are trained to generate text, not to make precise, interpretable numeric forecasts required for investing ([07:54]).
• ML models for investing (e.g., regression/decision trees) are better suited for stable, interpretable, and numerical predictions. Interpretability is key for regulatory comfort and internal understanding.
  • Quote: “They’re very stable approaches...LLMs are not particularly stable...and they’re very hard to interpret.” —David ([07:54])

3. How Decision Trees Work in Practice

• Decision Trees split data at nodes based on specific signals/features. Each branch represents a “if-then” condition, assigning stocks to buckets that guide forecasted return ([10:36]).
• Many decision trees are combined (e.g., via “gradient boosting”) to improve model accuracy and capture complex interactions.

4. Data Choices: Traditional vs. Alternative Data

• Traditional Data: Price history, fundamental metrics, analyst forecasts, and positioning; offers long time series and complete coverage.
• Alternative Data: Examples include satellite imagery, social media, smartphone data—exciting, but typically lacks breadth/depth for robust ML model training ([13:42]).
  • Quote: “Using more traditional data, you’re generally going to have a feature score for every company in your universe...If we start incorporating more alternative data, we don’t have the history, we don’t have the breadth. It creates a lot of challenges.” —David ([13:42])

5. Balancing Rationality and Empiricism in Feature Selection

• Debate: Should all model features have a rational link to economics/behavior, or can "black box" features with no obvious tie be used if they work?
• Pictet’s stance: Preference for rational/trusted features to start, but willing to let the machine explore combinations and accept that many discovered relationships may defy human intuition ([16:30], [18:49]).
  • Quote: “We like rationale in our inputs, but increasingly what the machine finds in the relationships between those different pieces, we do just have to accept that they’re a little bit less obvious.” —David ([18:49])

6. Optimal Time Horizons

• Shorter Horizons: ML adds more value in forecasting next-month returns or shorter; as the time frame extends, traditional factors (value, quality, etc.) become relatively more predictive ([21:00]).
  • Quote: “The shorter the horizon becomes, the more beneficial the machine learning element becomes within it.” —David ([21:00])
• Why? Short-term returns are driven by a mix of noisy, complex factors—ML is better at decoding this than linear models ([22:25]).

7. Improving Traditional Factor Strategies

• ML can enhance traditional factors (value, momentum, etc.) if properly applied, but the techniques/tools should match the task.
• ML approaches (e.g., LLMs) might be used for better news or transcript analysis, but aren’t suitable for every step ([23:46]).

8. Addressing the Critique: Is ML Just Data Mining?

• ML does "mine" data, but is safeguarded with clear guardrails against overfitting—using rationale in features, cross-validation, and testing for stability/the risk of decaying relationships ([25:52]).
  • Quote: “It kind of is data mining...but clearly you have to put a huge amount of guardrails in place to check for some of the challenges...” —David ([25:52])

9. Human vs. Machine: The Division of Labor

• Humans: Best at generating, curating, and selecting features; final oversight of portfolio construction and risk.
• Machines: Optimal for discovering which features/combos best forecast returns, structuring the model, and portfolio optimization ([28:11]).

10. Model Maintenance and Retraining

• Features and data sets are routinely tested and added; model is fully retrained every three months on a rolling 15-year window, ensuring robustness to regime changes ([32:53]).
• Efficiency has improved with tech, but core predictive relationships have shown surprising long-term stability ([34:49]).
  • Quote: “Versions of this model that we could train with 90s data have a lot of the same relationships between the features in it to now.” —David ([34:49])

11. Human Oversight Remains Crucial

• While automation has increased, clients want to see human oversight and supervision for trust, accountability, and risk management ([38:04], [41:10]).
  • Quote: “They do want to know that there is very strong human oversight, that there is guardrails and structure and supervision...” —David ([38:04])

12. ML vs. ChatGPT for Stock Picking

• LLMs like ChatGPT aren’t suitable for robust investment models due to lack of purpose-built training, inability to recreate time-restricted backtests, and lack of interpretability ([43:11]).
• Pictet uses interpretable, decision-tree models—every output can be traced to feature combinations, maintaining transparency ([43:11], [45:33]).

13. Portfolio Construction & Application

• Model is built for broad diversification—works as enhanced indexation (small tilts to alpha forecast) or as leveraged long-short for clients with more risk appetite ([47:53]).
• Controls for sector, factor, and region neutrality; alpha derived from stock-specific signals ([49:57]).

14. Cross-Market Transferability & Stability

• Relationships discovered by the model are stable and transferable across geographies (US, Europe, Emerging Markets), as long as the universe is broad enough ([52:37]).

15. Model Rebalancing & Trading Costs

• Model updates daily; portfolios are rebalanced weekly for efficiency and to control transaction costs. Human judgment decides rebalancing frequency, backed by extensive data analysis ([54:43], [64:04]).
  • Quote: “Trading costs and understanding the speed and horizon of your forecasts...is so incredibly important and so easy to just give up your alpha.” —David ([64:04])

16. Surprises in Feature Importance

• Many of the most predictive combinations are not obvious (“conditioning effects” and multi-signal interactions).
• Example: Calendar effects (proximity to earnings) combined with analyst upgrade/downgrade ratios and short interest played a stronger role than expected ([56:38], [60:32]).
  • Quote: “It’s a lot of those combinations of things working together. That is the power of the model.” —David ([60:32])

17. Closing Reflections

• On Edge:
  • David argues against focusing on narrow “edges”—it's about refining every stage, from feature engineering to training to portfolio construction ([63:12]).
• Advice for Investors:
  • Don’t overtrade. Even great alpha can be eaten up by excessive trading and costs ([64:04]).

Notable Quotes & Moments

[03:54] — Defining AI & ML

"AI, artificial intelligence, I simply think it means a machine doing some kind of function or role that a human would generally do..." —David

[07:54] — Stability and Interpretability

"They’re very stable approaches...LLMs are not particularly stable...and they’re very hard to interpret." —David

[13:42] — Data Breadth & Depth

"Using more traditional data, you’re generally going to have a feature score for every company in your universe...If we start incorporating more alternative data, we don’t have the history, we don’t have the breadth..." —David

[22:25] — Time Horizons

"Once you get down to the shorter horizon, again, sort of counterintuitively, lots of different things are driving returns. It gets quite noisy at that point...machine learning does give you that ability..." —David

[34:49] — Stability Over Decades

"Versions of this model that we could train with 90s data have a lot of the same relationships between the features in it to now." —David

[38:04] — Role of Human Oversight

"They do want to know that there is very strong human oversight, that there is guardrails and structure and supervision around how these models are trained..." —David

[43:11] — LLM Limitations

"I really remain very skeptical that [LLMs] are going to be able to do [investment strategy construction] effectively..." —David

[60:32] — Power of Feature Combinations

"It’s a lot of those combinations of things working together. That is the power of the model." —David

[64:04] — Investor Advice

"Don’t over trade...Trading costs and understanding the speed and horizon of your forecasts...is so incredibly important and so easy to just give up your alpha." —David

Timestamps for Key Segments

• [03:54] AI vs. Machine Learning in Investing
• [07:54] LLMs vs. ML Models for Stock Selection
• [13:42] Data Choices: Traditional vs. Alternative
• [18:49] How Much Rationality is Needed for Model Inputs?
• [21:00] ML’s Value at Short vs. Long Time Frames
• [25:52] ML: Data Mining or Not?
• [28:11] Human vs. Machine: Roles and Division of Labor
• [32:53] Ongoing Model Maintenance and Retraining
• [34:49] Surprising Stability of Relationships
• [38:04] The Necessity of Human Oversight
• [43:11] Limits of ChatGPT and LLMs in Investing
• [47:53] Model Application: Diversified, Enhanced Index & Long/Short
• [54:43] Model Rebalancing and Human Decision-Making
• [56:38], [60:32] Surprising Conditioning Effects and Feature Interactions
• [63:12] On Edge and Competition
• [64:04] Closing Investing Advice: Don’t Overtrade

Conclusion

David Wright’s perspective underscores that while machine learning offers a powerful edge in parsing noisy data and uncovering multi-factor relationships—particularly over short horizons—human oversight remains vital in feature selection and risk management. The future is likely to see increasing automation and broader model application, but with strong human guardrails, an emphasis on interpretability, and a continued focus on trading discipline.