The Inferential: What to Expect

The Inferential will be publishing original, considered writing about practical data science and machine learning topics, aimed at practitioners.

• On asking the right questions. Asking the right questions and correctly framing the problem is the most important part of any project — and the root of most failures. It's always worth taking the time to ask: Am I asking the right questions? Doing good applied work means thinking critically about trade-offs, assumptions, edge cases, failure modes, and misconceptions — and that starts with asking good questions.
• On evidence and uncertainty. I believe in updating beliefs based on new evidence. Principles are important, but so is being flexible enough to change your mind. I'm skeptical of big claims that don't acknowledge uncertainty. Predicting things is hard, and we often don't understand how systems work as well as we think we do.
• On intuition. I think developing good intuition is more important (and harder) than developing good coding skills. (If you don't understand why you're doing what you're doing, it doesn't matter how well you can code.)
• On domain expertise. I think good theory and domain expertise are prerequisites for good applied machine learning work. You need domain expertise to ask the right questions.
• On choosing methods. I think using the most appropriate method, after weighing the trade-offs, is better than using the newest one simply because it's the newest one. Always use the most appropriate tool for the problem, not the most impressive one.
• On good data. I think the quality of your data is more important than the sophistication of your model architecture. The same goes for your measurement strategy. Also, I think better data solves more problems than more compute.
• On preferences, institutions, and incentives. I'm interested in how institutional design shapes incentive structures. I like thinking about what rational, strategic behavior looks like, conditional on people's preferences (which are socially constructed) — and then thinking about how and why real-world behavior deviates from that baseline. (I'm a game theorist by training.)
• On interdisciplinarity. I think interdisciplinarity and cross-field exchange is valuable. (In particular, there's a lot that data scientists and machine learning engineers can learn from quantitative social scientists.)
• On communication. I like clear, concise writing. I think it's important to communicate technical material concisely and plainly. It's fine to use technical language when you need to be precise, but it's important to clearly communicate the intuition. Also, I like good data visualization and good data storytelling.

• Practical decision-making. When should you fine-tune a specialized BERT model versus prompting a frontier LLM? Are your evaluation metrics actually aligned with your business objectives? I'm going to write about how to evaluate trade-offs, identify assumptions, handle edge cases, and resolve failure modes.
• Research design and causal inference. How do you know that adding HyDE to your recommender system has improved performance? How do you test a new feature's impact on customer behavior when a clean, experimental A/B test isn't logistically viable? I'm going to write about how you can apply best practices from research design and causal inference to evaluate and improve production systems.
• Legal tech. Drawing from my experience building databases and training machine learning models around the Court of Justice of the European Union (CJEU) and European Union (EU) competition law, I’m going to write about how to train and deploy production-grade models in technical, specialized domains, like legal tech.
• Original analysis. From time to time, I'm going to publish concise, rigorous analyses that apply a wide variety of empirical methods. Some of these will take on serious business or social questions, and others will just be for fun. I'll use these to make substantive points and to illustrate how to use interesting methods.