Check Github and Bitbucket to see what I’ve been up to. Otherwise, here’s a quick roundup…

`hsplus`

is a Python library (with R bindings) that provides estimates for quantities involving the Horseshoe and Horseshoe+ shrinkage priors.

It also contains general numerical estimation procedures for bivariate confluent hypergeometric functions, as well as symbolic SymPy implementations.

`amimodels`

is a Python library that provides core implementations of models designed for use with Advanced Metering Infrastructure (AMI) data in `eemeter`

. The implementations are fundamentally Bayesian state-space and mixture models that automatically account for the systematic changes, missing data and varied observation frequencies. The models and custom MCMC estimation methods are written in PyMC2 and—as such—are easily extensible.

Complete R and Stan Horseshoe and Horseshoe+ shrinkage prior examples from our paper “Default Bayesian Analysis with Horseshoe Estimators”:

- sparse multivariate normal estimation
- ratio of normals estimation
- product of normals estimation
- max of multivariate normal estimation
- sum of squares estimation

Bus Time is the open source Java suite that provides real-time bus tracking to NYC. I designed and developed the statistical inference capabilities and helped build the production service components. The model handles free and constrained location tracking along street networks, inference for unobserved operational states (e.g. in layover, at a stop, in progress) and path-based states (e.g. current trip, route, run), as well as inference for faulty operator input (e.g. operator ids, sign codes).

In production the model handles real-time updates at ~30 second intervals for hundreds of routes and thousands of buses simultaneously. Its statistical specification is Bayesian and its estimation is performed by a custom particle filter.

`prox-methods`

is a very experimental R package with C++ implementations (via Rcpp) for some of the proximal optimization methods from the paper “Proximal Algorithms in Statistics and Machine Learning”.

`open-tracking-tools`

is an open-source vehicle tracking library that implements custom Particle Filters to infer locations, paths and on/off-road states. Given a transit graph, `open-tracking-tools`

provides robust real-time Bayesian inference for noisy GPS data. OpenTripPlanner graph support is built in, so street information encoded in OpenStreetMap can be used with fairly minimal effort.

Extensions to the Cognitive Foundry API including, but not limited to, specialized distributions, sampling techniques, and numerically stable computations for Dynamic Linear Models.

`ParticleBayes`

is an R package implementing a collection of particle filters for hierarchical Bayesian models that perform sequential parameter estimation.

Java code for Bayesian models that are estimated by Particle Filters and implement parameter learning.

Big data simulation of Chicago’s public transportation to improve transit planning and reduce bus crowding.