// signal log
I follow things until I understand them, then write down what I found. Subject matter varies wildly.
Price per hundredweight falls as cattle get heavier, which reads as "heavier is worth less." It is not. Chasing that down cost me three assumptions and produced one number I did not expect: this market pays about half again as much for a pound added to a mature animal as to a growing one, and it has done so in every month of the record.
Steer prices at one Alabama sale barn tripled in seven years, and the model dates every break. The interesting one happened eleven months earlier, in a series nobody watches.
Expected batting average is decided at the moment of contact, so it never learns where the ball went or how far it carried. T-Mobile Park lives in that gap — and the marine layer everyone blames for it turns out to be a null result.
I tried to put a Bayesian changepoint analysis of Nashville 311 into Metabase. Two panels, a credible interval, and a probability bar chart — it took 659 lines of Python against an HTTP API, and I still could not review it. Lightdash gets this right by making the dashboard a file.
I ran a Bayesian changepoint model over nine years of Nashville 311 data to find the moment my old neighbourhood changed. There was no moment.
Astro on Cloudflare Workers — static pages, on-demand API routes, one deploy.
An instrument for reading what expected batting average cannot see. Pick a contact cell and watch its outcome spread across the field, in any of the thirty ballparks.
Fantasy football analytics for a Sleeper league — power rankings, all-play records, trade grading, and a scoring system for bad lineup decisions.
Seven years of USDA sale-barn reports for one Alabama cattle auction — Bayesian market regimes, a time-varying value of gain, and a validation rule that caught a typo in the source data.