Probabilistic temperature guidance from NBM
Hello all. I'm not a Redditor, but I've been aware of Kalshi's weather prediction markets for some time. Reading your posts/comments about them, I noticed some recurring struggles and did a little Saturday project that I thought might be useful.
For my own professional reasons, I use various NWS forecast products extensively, including the National Blend of Models (NBM). If you are not familiar, the NBM aggregates a large group of NWP forecast products from the American weather center and others and applies a bias correction based on how all the models performed over recent history. The details are genuinely complicated, but it's a good product and there's a ton of information out there, explaining how it works, if you're interested. It's widely considered the best forecast for the US (for most purposes) and is almost always the benchmark that other US forecasts are compared to.
Importantly, the NBM project has a probabilistic subcomponent called Quantile Mapping and Dressing (QMD) that forecasts common surface variables in percentile terms. I thought the NBM's QMD might be helpful for you guys to get a sense of forecast uncertainty, so Claude and I put up a little app here: https://pointnbm.com/
You are also free to pull the JSON for different stations directly and use it however you want, like so: https://pointnbm.com/data/KLAX/latest.json
NOTE: I'm not selling anything. It's annoying to deal with the underlying gridded data directly, but it already gets extracted/processed on my own infra for other reasons. Cloudflare serves and caches the point data for the app, which is very light for individual locations, so it's nearly free to host this even if it gets traffic.