Episode
12

Stefan Wager on Why Your Machine Learning Solves The Wrong Problem

Stefan Wager—Professor at Stanford and expert on causal machine learning—has worked with leading tech companies including Dropbox, Facebook, Google, and Uber. He challenges the widespread assumption that better predictions mean better decisions. Traditional machine learning excels at prediction, but is prediction really what your business needs? Stefan explores why predictive models alone often fail to answer critical “what-if” questions, how causal machine learning bridges this gap, and provides practical advice for how you can start applying causal ML at work.
March 17, 2025
Listen on
spotify logoApple podcast logo
Guest
Stefan Wager

Stanford University

,
Stefan Wager is an associate professor of operations, information and technology at Stanford Graduate School of Business, and an associate professor of statistics (by courtesy). He received his PhD in statistics from Stanford University in 2016, and also holds a BS (2011) degree in mathematics from Stanford. He was a postdoctoral researcher at Columbia University during the academic year 2016–2017, and has worked with or consulted for several Silicon Valley companies, including Dropbox, Facebook, Google, and Uber.
HOST
Hugo Bowne-Anderson

Delphina

Hugo Bowne-Anderson is an independent data and AI consultant with extensive experience in the tech industry. He is the host of the industry podcast Vanishing Gradients, a podcast exploring developments in data science and AI. Previously, Hugo served as Head of Developer Relations at Outerbounds and held roles at Coiled and DataCamp, where his work in data science education reached over 3 million learners. He has taught at Yale University, Cold Spring Harbor Laboratory, and conferences like SciPy and PyCon, and is a passionate advocate for democratizing data skills and open-source tools.

Key Quotes

Key Takeaways

1. Machine Learning Predicts, but Decisions Require Causation

Classical ML excels at predicting the status quo—but most organizations aren’t trying to predict the world, they’re trying to change it. Stefan explains why prediction alone fails at critical “what-if?” questions.

2. Beware of Churn Prediction Pitfalls—Use Experiments to Find Real Impact

Predicting churn probability alone won’t help you choose effective interventions. Stefan shares concrete examples from industry—like loyalty programs—where predictive models have led companies astray, and shows how causal ML corrects these mistakes.

3. Experiments Are Crucial—and Easy to Get Wrong

Causal ML depends heavily on experimental data. Stefan highlights common pitfalls in running experiments, such as accidental biases that invalidate results. Getting experimentation right is fundamental to success.

4. Start with Clarity—Define Actions, Not Just Predictions

The hardest part of causal ML isn’t running models, but clearly defining the decisions you want to make. Stefan emphasizes that actionable questions must drive data collection, modeling choices, and interpretation.

5. Causal ML Requires Rethinking the ML Workflow

To effectively deploy causal ML, you need to move beyond traditional “XY” prediction workflows and integrate explicit causal reasoning into data collection, model development, and business strategy.

5. Tools Matter, but Thinking Matters More

Stefan suggests mastering just a few powerful causal ML tools—like generalized random forests—rather than chasing every new method. But he emphasizes that good causal inference always starts with careful, strategic thinking about your business problem.

You can read the full transcript here.

00:00 The Limitations of Prediction

01:08 Causal Machine Learning: A New Approach

04:20 Introducing Stefan Wager

04:23 The Importance of Causal Inference

07:31 Challenges and Adoption in Industry

15:52 Practical Examples and Case Studies

20:25 Implementing Causal ML in Organizations

25:14 The Value of Experiments in Causal Analysis

25:34 Challenges with Observational Data

26:12 Industry's Approach to Causal Inference

27:11 Historical Examples and Model Evaluation

28:30 Heuristics for Choosing Modeling Techniques

28:56 Tree-Based Methods and GRF Software

33:27 Communicating Causal ML Results

37:00 Explainable ML vs. Causal ML

40:37 Causal Discovery in Different Fields

42:44 Failure Modes in Causal ML

45:03 Industry vs. Academia in Causal ML

49:18 Resources for Learning Causal Inference

50:43 Future of Causal ML in Business

52:08 Final Thoughts and Common Sense in ML

Transcript

featured

In the spotlight: Our most popular episodes

most recent

Listen up: Our latest discussions

Hear the hottest takes on data science and AI.

Get the latest episodes in your inbox

Never miss an episode of High Signal by signing up for the Delphina newsletter.

By clicking Sign Up you're confirming that you agree with our Terms and Conditions.
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.