I'm a Bayesian statistician with 20+ years of experience in academics and industry. Much of my work has been as a data science enabler, by building tools that data scientists can use in their analyses. I'm an expert in Bayesian computation, specializing in applied problems using Markov chain Monte Carlo. I have extensive experience with hidden Markov models, time series modeling, mixture models, hierarchical models, and various forms of Bayesian regression.
I have held director-level positions at Google and Capital One, and served on the faculty at USC's Marshall School of Business.
Multi-armed bandit experiments
Multi-armed bandit experiments are a way of conducting A/B tests (and more general sequential experiments) with much lower opportunity cost than a traditional statistical experiment.
My work on multi-armed bandits became the basis for Content Experiments by Google Analytics, and found its way into parts of the ad system as well.
For an overview of the Thompson sampling method for multi-armed bandits, see the papers and slides below.
The talk below summarizes my work on Monte Carlo methods for big data sets in a MapReduce environment.
I am the author of the bsts package for R, which fits Bayesian structural time series models to time series data. You can read about bsts on the Unofficial Google Data Science Blog, or for a more in depth view read my papers with Hal Varian.
- Predicting the present with Bayesian structural time series.
- Bayesian variable selection for nowcasting economic time series.
If you prefer slides, try these: