Nathan Walker

Nathan Walker

@bbstats.bsky.social

currently: data science / optical @ NFL prev: data science @ GSW sometimes bball analytics openRAPM: https://openRAPM.com/ chimeraboost: https://bbstats.github.io/chimeraboost/

2.7K followers1.5K following313 posts

Nathan Walker reposted

Kate Compton
@galaxykate.bsky.social

You can test new tech ideas using the Seinfeld Test Would the product eliminate the plot of an episode? (Google maps, cell phones, paypal, battery packs) Good tech. Would the product inspire new Seinfeld plots? (NFTs, AI chatbots, crypto currency, blindboxes, metaverse land sales) Bad tech.

Nathan Walker
@bbstats.bsky.social

3. The simplest solution is to do a full refit! Something I have done manually in the past. But this roughly doubles the training time of the model What if there was some way to retain most of the tree structure your GBDT, but increase the data it sees to the full dataset?

Nathan Walker
@bbstats.bsky.social

Introducing 'replay'! github.com/bbstats/chim... 1. Fit with early stopping 2. Using the same trees & splits (skipping ~85% of train time), re-calculate the leaf values (& other data-dependent parameters). In my testing, 10-40% speed improvement over a 'full-refit' and the speed scales with size!

github.comReuse the tree structures when refitting on the full data: same accuracy, a third less fit time by bbstats · Pull Request #41 · bbstats/chimeraboostWhat changes: refit_full now defaults to "replay". Same accuracy, roughly a third less time to fit. Why we currently fit twice With early_stopping=True (the default), ChimeraBoost holds o...
Nathan Walker
@bbstats.bsky.social

Doing the minimal possible SEO here but - I (and Claude) developed a python-only (e.g. numba!) backend Gradient Boosting library that is insanely fast on small/medium data (eg your prototyping phase!) pip install it and tell me what you think github.com/bbstats/chim...

github.comGitHub - bbstats/chimeraboost: A Python-backend gradient boosting libraryA Python-backend gradient boosting library. Contribute to bbstats/chimeraboost development by creating an account on GitHub.

Nathan Walker reposted

Dave DuFour 🐉
@davedufour.bsky.social

Victor Wembanyama is basketball Genghis Khan

Patrick Wyman@patrickwyman.bsky.social

This is the most excited I've been about an NBA Finals in a while - so happy for the Knicks and their long-suffering fans, and if you're not stoked about Wemby and the Spurs, I don't think you actually like basketball

Nathan Walker
@bbstats.bsky.social

Caveat A: Specialists for specialist scenarios (single defensive possession stoppers in close games) probably have more utility there for example Caveat B: "Over the course of their contract" of course depends on how tradeable the contract is / becomes.

Nathan Walker
@bbstats.bsky.social

GPT 5.1 1455 Arena Elo (current) November 12, 2025 $7.81 per million weighted tokens Gemma 4 1451 Arena Elo (current) April 2, 2026 $0.34 per million weighted tokens