Robert's BayesCamp story

Case studies

I started BayesCamp in 2017 to provide training and coaching for people who analyse data. We aim to do what the big providers neglect:

• supporting managers, recruiters and others who work with data analysis teams, who want to know more, but don't need to be experts themselves
• small-group, flexible, tailored learning
• one-to-one coaching
• primers in more advanced topics like Bayesian modelling, effective data visualisation, and social network analysis
• all trainers have been there and done it

But there's something we don't do: we don't give you any hype. When some software or algorithm doesn't meet expectations, we'll tell you so.

I've worked on hospital quality and safety data, evidence synthesis, a wide range of health, social care and education projects, teaching research methods and data analysis to people from all over the world, and consulting for clients in private, public and charitable sectors.

Have you ever sat in a training course where everybody was kept busy with post-it notes and word association, but nobody had learnt anything by the end of the day? I certainly have. I realised that employees deserve learning experiences as good as any university course.

There's no reason why it should be dumbed down just because it lasts one or two days. And the trainer should be someone who has worked on serious projects, not just pressing buttons but also communicating results and thinking about the interpersonal skills that are needed.

The Economist's EAGLE golf system

The Economist's sports blog had teamed up with Dell to take a collection of R programs that ran multiple projections of how golf tournaments would play out, and make them into a system that would update with live data, and feed the predictions into a website.

Data Editor, Dan Rosenheck, had spent three years developing the core of the system up to this point, and one of his goals was to push up the predictive ability of the system by giving it more flexibility. The code that took data from each player, each course, and each hole, and detected how player talent was evolving over time, could only generate future projections from certain formulas.

Dan knew that if he gave the system more freedom, it could pick up anomalies and give more accurate predictions. But the expert advice he got was that it just couldn't be done -- unless he could switch the core of the prediction code to Stan, Bayesian inference software with a reputation for being difficult to learn and use. And meanwhile, the start of the 2019 majors season was getting ever closer!

That's when he got in touch with BayesCamp. We were able to sit down with him and take the time to understand his code, the constraints he was working in, and the output format that the front-end developers would need. We wrote some new functions within Stan for the weird distribution that golf scores have, and rebuilt the core of his probabilistic model to run in Stan's Hamiltonian Monte Carlo algorithm.

The new system updates the model with Stan and uses parallel computing in the cloud to get that done in the timescales required. We got it up and running just in time for the launch... and just in time for Tiger Woods' freakish comeback to catch out every prediction out there. Well, the most likely prediction isn't going to happen every time -- that's probability.

You can check out Dan's system, EAGLE, while golf majors are in progress at eagle.economist.com