What this covers
Michael Reece walks through how he builds detailed, continuously updated models of consumer businesses from transaction-level data, with Patrick and Michael asking questions and offering analogies. The conversation opens with the idea that conservative investment firms keep repeating what has worked, which leaves room for an orthogonal disrupter in the pattern of Clayton Christensen. It then moves to what the digital residue of online activity and cheap electronics reveals about who is winning in a market, including customer attributes a company never disclosed.
The concrete examples carry most of the range. Blue Apron's IPO documents showed growth, but cohort decomposition by join date showed newer cohorts were less sticky and costlier to acquire. Walmart's best 9 percent of customers generate half its revenue, and a Pareto concentration trajectory over time is read as a sign of business health. New-store sales are time-shifted to a common opening date, in the manner of honeymoon-period normalization, to isolate organic growth, cannibalization and maximum store density.
The contentious turns are about method. Reece argues for hypothesis-first analysis and calls running every algorithm and keeping the best score lazy data mining. He holds that machine learning suits stationary business growth but is a poor fit for price prediction, and that models find deterioration easier to predict than excellence. The talk also covers the quant-discretionary spectrum, falling S&P 500 tenure, spreadsheets versus Jupyter notebooks, the speed of change in online advertising, and why many independent explorers beat a single top-down plan.
Reece argues that alpha comes from processing raw transactional data into granular, business-specific models rather than from data access itself, and that machine learning works only on stationary business fundamentals, not on price prediction.
- Data is commoditized but information extracted from it is not; most firms only roll data up to top-line revenue
- Granular cohort, product, and geography models reveal business health that headline financials hide
- Machine learning is well suited to stationary business growth patterns and poorly suited to non-stationary price action
The data-investing edge comes from forming good hypotheses and combining business with statistical and ML knowledge.
- As answering questions with tools gets easier, the differentiating skill in data investing becomes forming good hypotheses, and the people who adopt new methods should cycle quickly through questions and experiments.
“what's becoming most important is the hypothesis, is asking the question. Answering is getting easier and easier and easier.”
- Building an information edge requires both a fundamental understanding of business and statistical and machine learning methods, which makes the task daunting but potentially rewarding.
“It requires both a fundamental understanding of business and of data statistics and methods like machine learning.”
Machine learning fails at predicting stock prices but works well for extracting and classifying data.
- In the host's own work, machine learning was found to be useless for predicting future stock prices but extremely useful for extracting and classifying data.
“In our own work, we found machine learning to be useless for predicting future stock prices, but extremely useful for other things like extracting and classifying data.”