Below is our first edition and we want to make it a newsletter, we will continue to post on discourse publicity. Please sign up at:
Stan this Month
Editor: Charles Margossian
ISSUE 1, JANUARY 2019
Stan this month is a newsletter aimed at highlighting major developments, discussions, and ideas within the Stan community. It provides an overview and pointers to more detailed references.
This month’s issue paints 2018 in broad strokes.
The Stan governing body
Our community of users and developers is growing, which means the Stan project needs to scale up. To do so, we formed a Stan governing body (SGB). The SGB’s mission is to oversee all aspects of the Stan project. This includes managing online resources, handling various funding sources, and supervising the organization of Stan conferences. The SGB is also strongly committed to advancing diversity in the Stan community.
More information about the members and the mission of the SGB can be found on https://mc-stan.org/about/.
Major software developments
In 2018, we dove deep into the high-performance arena. Some of our biggest new features include:
Within-chain parallelization . The parallel map function allows users to parallelize the computation of distributions and their gradients. The performance scales up (roughly) linearly with the number of cores.
Accelerated compound GLM functions. Allows users to tackle more efficiently generalized linear models.
Differential algebraic equation (DAE) solver. Solves systems of equations that couple ordinary differential and algebraic equations. It is one of the many features being developed to support differential equation based models.
Other improvements include new features, bug fixes, as well as drastic improvements made to the parser and the language itself. There is in addition a growing ecosystem of softwares build around Stan for more specialized applications, such as the popular R packages RStanArm and BRMS, or the Python package Prophet developed by Facebook for forecasting.
Contributions to the Stan Conferences provide an extensive list of applications, and can be found on https://github.com/stan-dev/stancontalks. These contributions span a broad range of topics, including physics, genetics, and social sciences; with models using hierarchical structures, predictive time series, and partial differential equations.
Articles using Stan can also be found in various venues. Our editorial highlight include a paper on industrial optimization and one on the development of a new method and its application to pharmacometrics:
- Scaling Auctions as Insurance: A Case Study in Infrastructure Procuremen t by (Bolotnyy & Vasserman, 2018) [link]
- Bayesian aggregation of average data: An application in drug development by (Weber et al., 2018) [link]
Theory and method
Concurrent with software development and application (and often overlapping with), a lot of research is done at a more theoretical and methodological level. Papers on these subjects either give users practical guidance on how to tackle difficult problems or lay the foundation for long-term developments. These are our editorial highlights from last year:
- Visualization in Bayesian workflow by (Garby, Simpson, Vehtari, Betancourt, & Gelman, 2018) [link]
- Validating Bayesian Inference Algorithms with Simulation-Based Calibration by (Talts, Betancourt, Simpson, Vehtari, & Gelman, 2018). [link]
- Yes, but Did It Work? Evaluating Variational Inference by (Yao, Vehtari, Simpson, & Gelman, 2018b) [link]
- Using stacking to average Bayesian predictive distributions (with discussion) by (Yao, Vehtari, Simpson, & Gelman, 2018a) [link]
- Geometric Theory of Higher-Order Automatic Differentiation by (Betancourt, 2018) [link]
Events and community
We have a strong community of users and developers, eager to help one another and dedicated to open science. Discussion on the Stan forum and our GitHub pages are still going strong.
Rather exceptionally, we had two dedicated Stan conferences this year: one in Asilomar, California in January, and one in Helsinki, Finland in August. Each event hosted about 200 participants. Talks were recorded and often accompanied by a notebook with computer code. This material is now freely available online. On top that, we offered tutorials covering introductory Stan, advanced Bayesian modeling, and how to develop Stan features at a C++ level.
There have been many more events, such as the Stan for Pharmacometrics day in Paris, France; the many events organized by meetup groups in New York, Boston, Berlin, and now in South Korea; and many workshops and courses hosted by various conferences and institutions.
A big part of building Stan and its community goes through the development of pedagogical tools. The development team is actively working on the Stan book , which is based on the Stan User Manual . It is still a work in progress, but the draft looks fantastic.
Michael Betancourt has also been prolific at writing case studies, often available both in R and in Python. Our editorial highlight is his case study entitled Towards a Principled Bayesian Workflow [link].
New Year’s resolutions
It is January after all. Ultimately, our efforts will continue to be driven by the community. One suggestion for 2019 would be to get involved in answering the increasing volume of questions on the Stan forum, review more pull requests if you are a developer, diligently report bugs and issues, and – last but not least – continue sharing your work. Tell us what your achievements and what your challenges are.
Good venues to do all these are the Stan Forum and our GitHub pages. You may also contact the editors of Stan this month via email@example.com.
The Stan development team
Editor: Charles Margossian
Photo Credit (CC-BY licence) for Audience at Stan Con Helsinki : Aki Vehtari