R-MAP Bracket

BRACKET answers the question every forecast avoids: how wrong is this likely to be? Point forecasts are everywhere. Intervals are rare, and the ones that exist usually come from a model's own assumptions — which fail precisely when the world stops behaving like the model. BRACKET calibrates on your own recent misses instead. The result carries a distribution-free coverage guarantee: at the tolerance you choose, the truth lands inside the bracket at that rate, whether the forecast came from a spreadsheet, a vendor model, or your gut. THE BRACKET Your next forecast, and the interval around it. Split conformal prediction on the finite-sample quantile — so when there are not yet enough resolved points to back the level you asked for, BRACKET says how many more it needs instead of quoting an interval it cannot defend. THE RECORD You asked for 90%. Did you get it? Every resolved point is scored against the bracket BRACKET would have quoted before it resolved — no hindsight — and the achieved rate carries the confidence interval a count that size deserves. THE WATCH A coverage guarantee assumes your misses stay exchangeable, and that is exactly what breaks when a process shifts. BRACKET watches for it: max-CUSUM against a seeded permutation null, plus BOCPD, ADWIN, PELT and SPC rule packs on one board. When a shift is found the calibration window truncates at it and the app says out loud that the guarantee is now weaker. The null distribution is drawn, so you see the shape your p-value was read off rather than only the number. THE LAB, AND THE REPLAY Sweep the knobs the Watch holds fixed — hazard rate, ADWIN delta, PELT penalty — and watch the coverage surface move. Then run every method against the same series and watch them disagree: where two of them part company is the most useful thing a change-point tool can show you. WATCHED SERIES Not everything is forecast. For numbers you simply track — defect rate, churn, cycle time, spend — BRACKET runs Mann-Kendall with Sen's slope and the same change-point watch, and gives one plain-language verdict. AT A GLANCE The Board puts every series in one view: how many you watch, how many the watch has flagged, which ones moved. A configurable widget — pick a series, or leave it on the next bracket due. The lock-screen circular face draws achieved coverage as an arc with your claimed level as a tick on the same arc, so "am I under my own promise" reads without a number. Long computations report a true percentage of real work. BUILT TO BE CHECKED Every p-value comes from a seed you can see and change, and every verdict quotes that seed and its permutation count. Re-test one to watch the p-value wobble within its own sampling error — the honest uncertainty in the uncertainty. Press and hold any chart to read a value off it; tap any headline figure to find out what it means. WORKS WITH THE R-MAP SUITE A bracket is a three-point range, so BRACKET exports rmap.portfolio.v1 — the format KEYSTONE reads as a severity range. A measured interval replaces a guessed range. PRIVACY No accounts, no feeds, no backend. Everything is computed and stored on your device, and nothing is transmitted anywhere. FREE AND PAID Three tabs are free permanently — Series, Watch and Settings: the interval and its guarantee, the change-point watch, the drawn null, the trend verdict and the scorecard, on three worked-example packs plus one forecasted and one watched series of your own. A single purchase, not a subscription, unlocks the three paid tabs — the Board, the Lab and the Replay with its method race — and then, on your own series: unlimited series, the coverage ledger, the Watch across series, the widget, the report and the exchange file. Demonstration data is illustrative and labeled. BRACKET forecasts nothing itself — it brackets the forecasts you supply — and is an analytical instrument, not financial or investment advice.