Articles, newest first
Temporal Validation in Machine Learning: Testing Models Against the Future
Temporal validation evaluates machine learning models the way they will be used: trained on the past and tested on the future.
Read article
CUPED and Regression Adjustment: Cutting A/B Test Variance With Data You Already Have
The experiment needs 1,570 users per arm to detect a two percent lift. Each user's spend over the previous month is already in the warehouse and correlates 0...
Read article
Randomness Does Not Owe Us a Reversal
Four heads in a row do not make tails due on an independent fair coin. But history can matter in other mechanisms. Three examples show exactly when the next ...
Read article
Novelty and Primacy: When the Effect You Measure Depends on How Long You Looked
The test reports a 9 percent lift after three days, 5.7 after two weeks and 3.9 after four. None of those is the answer. The effect on a user who has lived w...
Read article
How Big Does a Test Set Need to Be?
A challenger beats the incumbent by 0.8 accuracy points on 400 test cases. The standard error of that measurement is 1.5 points. The comparison was decided b...
Read article
Your Biological Age Is Not a Single Number
Epigenetic clocks are useful research tools, but “biological age” is not one directly observed property of the body. Different clocks estimate different cons...
Read article
Competing Risks, Recurrent Events, and Multi-State Models Are Not the Same Problem
A single event time is only one possible event-history object. When causes compete, events recur, subjects move through intermediate states, or a longitudina...
Read article
The Dangerous Push Toward Practical Mathematics: Why Pure Research Must Remain Protected
The pressure to justify mathematics by immediate usefulness misunderstands how mathematical progress works. Pure research needs protection precisely because ...
Read article
The Intellectual Crisis: How Utilitarian Thinking is Destroying the Foundation of Human Progress
Short-term thinking and utilitarian pressures are undermining the very institutions that sustain human progress. This is a crisis not just in science policy,...
Read article
Stability Is Not Truth
A continuous Gaussian population can produce highly stable clusters. A negative control, positive control and representation perturbation show what bootstrap...
Read article