top of page
Joe Marion

Hi, I'm Joe

I received my PhD in Statistical Science at Duke University in the summer of 2018. My research, supervised by Scott Schmidler, focused on theoretical properties of sequential Monte Carlo samplers. I'm currently designing complex and innovative clinical trials at Berry Consultants

EXPERIENCE
DATA SCIENCE
EXPERIENCE
Summer 2017

Data Science Intern

MAXPOINT

I improved existing methods designed to asses how online ad campaigns drive foot traffic to brick and mortar stores. This problem was challenging because most of the data is unobserved and the data we did observe was very noisy! I combined causal inference methods, outlier detection, and Bayesian modeling to address these problems.

Summer 2016

Research Intern

GEOMETRIC DATA ANALYTICS

I built anomaly detection models for vehicle traffic, using scraped Waze and transit system data. These models were deployed in real time during the 2016 Summer Olympics in Rio to help US agencies monitor the games. Having my models supporting real world events is incredibly satisfying.

Summer 2015

Data Science Intern

BLIZZARD ENTERTAINMENT

I designed an ensemble of machine learning algorithms to predict user churn, improving performance substantially. I then used causal inference methods to identify the reasons that users stopped playing games. Blizzard implemented many of my suggestions and you can still see them in Heroes of the Storm!

RESEARCH
PHD RESEARCH

I'm interested in Monte Carlo and simulation methods used for inference in Bayesian statistics problems. My research focuses on one sampling technique, sequential Monte Carlo (SMC).  This method isn't particularly new, but it hasn't been widely adopted in Bayesian statistics. 

​

I'm developing finite sample theory for SMC. Most of the current results in the field are asymptotic. While these results are valuable, I think it's important to have error bounds for the methods we use in practice (not just their infinite particle approximates). The approach I've taken to solving this problem is novel and it allows for  the explicit interrogation of the SMC algorithm.  

 

​

My results show that the complexity of SMC is roughly comparable to Markov chain Monte Carlo for many problems. I also show how selecting good sequences of SMC distributions (known as paths) can lead to SMC algorithms with superior complexity. Finally, I provide a convergence result for the commonly used adaptive step size approach to selecting SMC distributions.

​

My research is currently funded by the SAMSI program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applied Mathematics

ABOUT
ABOUT

Before coming to Duke, I studied Economics and Math at Cornell University. After graduating, I was commissioned as a Field Artillery Officer in the US Army. I deployed to Afghanistan and Iraq before returning to school to pursue my PhD in statistics.

​

I enjoy climbing (mostly bouldering), board games, and reading science fiction. Every year I take time to go hunting, usually for elk but sometimes for deer or pheasants. I enjoy traveling and playing board games with my wife Brianna.  In the wee hours of the night I paint miniatures, currently from Warhammer 40k.  

Joe Marion and Brianna Shaver
RESUME
CONTACT
bottom of page