🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Martin Ingram James Hensman This is a compact and clean post! Motivated by it, I put together a practical JAX implementation from scratch patel-zeel.github.io/blog/gp/2022/1…. Many neat tricks I learned during #GSoC2022 helped in making this work.

Does anyone know of a benchmark comparing JAX against TF for autodiffing big loops like in state space models? I’m pretty sure I saw a big speed up when using scan in JAX vs TF but don’t have the numbers to back up the claim…

Online linear regression is a cool special case of Kalman Filtering. Does anyone know how to _delete_ an observation? I'm interested in fitting a rolling window and it would be more efficient to do one update and one deletion rather than refitting each time...

Does anyone know a good soccer player rating? I'd be curious how players in the world cup are rated

The other day, I tweeted asking about how to forget an observation in an online linear regression. I got useful advice, and got it to work, so I thought I'd share the result in case it's of use to others. Here's the post: martiningram.github.io/forgetting/

Does anyone know of a nice python library that does a Bayesian logistic regression with a ridge penalty? E.g. Laplace approx. to marginalise, then MAP on the penalty term. Currently I'm using glmnet but I feel like this approach would be faster/better...?

Anyone have a favourite paper that introduces partial pooling? I’d like to present if at a reading group and am wondering what the best resource might be; maybe a BDA3 chapter?

I am excited to announce that the On The T blog is joining the 'You Cannot Be Serious Stats' team on Substack. If you are looking for new tennis analysis, check us out and sign up to receive new stories here seriousstats.substack.com

A new 'You Cannot Be Serious Stats' analysis looks at the players who likely benefited the most from the unusual absences of several top players at slams in 2022, including an AO and US Open without Djokovic seriousstats.substack.com/p/2023-01-10-s…

I'm contributing to this project, and am also excited about it!

Being a Bayesian, I like the derivation of Kalman Filters in terms of conjugate normals best. Somebody recently pointed out the fading memory Kalman Filter to me and I was wondering if it had a similar probabilistic interpretation; does anyone know?

What makes Novak Djokovic so lethal on the return? It all starts with his first shot. See the full breakdown of Djokovic's return effectiveness in the latest post on 'You Cannot Be Serious Stats' seriousstats.substack.com/p/2023-02-04-d…

I have never written a VB paper using SGD --- instead, I have always fixed the draws in advance and optimized the deterministic objective. Thanks in part to amazing co-first-author Martin Ingram, there is finally a paper about it! arxiv.org/abs/2304.05527

Anyone have ideas for avoiding variances in state space models from growing too large? E.g. where one of the states is very rarely observed, and the random walk just keeps increasing its variance. I've been capping them to the prior variance, which works OK but feels a bit hacky.

What are people's favourite ways of fitting an inducing point GP to big Gaussian data these days? I think I have a good overview of the non Gaussian case but not quite as sure about the Gaussian one.

Hey all, if you wanted to fit a linear mixed model with a large number of random effects per unit, what’s your favourite fast software? Let’s say we have 15+ coefficients that should vary by group. I think lme4 and statsmodels get pretty slow eg.