Track Record
Every prediction on this site is published before the number exists, then graded in public against the first print. This page is the whole record — hits, misses, and all.
We aim for ~80%, not 100% — a forecaster who’s never wrong is using bands too wide to mean anything.
On the badge-driving indicators, how often our forecast implied the same status — ok, watch, elevated, or alert — that the actual print did. Market prices carry no badge and are excluded. Our edge over a naive “same as last time” guess is broken out by category below — all recomputable from the public ledger.
On the record now
By category
Recent grades
How this works
Each week, every indicator’s full history is backtested: we stand at hundreds of past dates, predict the next print using only data available then, and measure the miss. Whichever simple statistical model wins — and only if it beats the naive “same as last time” guess by a real margin — becomes that indicator’s forecaster. The ranges we publish are the ranges those backtests earned. Every model is explainable in one sentence; there is no black box.
Predictions are graded against the number as first published, then frozen — later data revisions are footnoted, never regraded, and a graded entry never changes. Because the ledger lives in this site’s public git history, the timestamps are independently verifiable: every prediction is committed before its print exists.
We forecast market prices too — the scoreboard, currencies, commodities — but hold them to a different standard. Next week’s price move is the one thing honest models can’t call, so for these we publish only the range history suggests around today’s level, never a directional bet, and we keep them out of the skill score above — beating a coin flip isn’t an edge. The bands are still graded, so you can see how honest they are. None of it is investment advice.
data/predictions/.