-
Hear when your Stripe card declines jump, from your own GitHub account
A free watch that reads payments, revenue and declines from your own Stripe key and messages you when one moves to a new level. Synthetic shops: 20 of 20 heard.
-
Seven calls, one question: a 26-second analysis that only needed one fit
A several-columns analysis in the change-point detector ran the same 3.6-second stale-feed fit seven times. A 16-entry memo cut it to 4.6 s, same output.
-
The cycle check that cost more than finding the steps
One periodogram check took 679 ms on an 8,609-day series. A one-period trig table cut the whole analysis from 1.56 s to 0.92 s, same answers.
-
Taking out the daily cycle did not make the alerts point at anomalies
On NAB's 47 real hourly series, removing the daily cycle left the change watch's alerts near chance. A drift model still points at anomalies best.
-
The map nobody read: a changepoint engine 32% faster with identical answers
A Bayesian changepoint engine built a map of its whole posterior each step and read one entry. Removing that churn: 32% faster, identical output.
-
Two sensors agree on noise too: a pre-registered test of 'alert when two columns agree'
On a real 8-sensor pump, waiting for two sensors to agree cut false alerts from 4.2 to 2.7 per 100 checks, still over budget, and missed 9 of 33 faults.
-
A change, or just clustered volatility? Searching the cuts with a GARCH model
A pre-registered GARCH search for change points in returns: a third of the false alarms of a block bootstrap, and on 33 years of the Dow it cuts nothing.
-
A results table that measured nothing, and what the real numbers say
A table in our changepoint detector's README came from a script whose likelihoods were all NaN. Fixed, the idea works better, but still not well enough to ship.
-
A watch for a level that wanders
The daily watch now has a setting for prices and stock levels that drift. Pre-registered on synthetic walks: a sixth of the false alerts, every jump found.
-
The 1976 warming acceleration, dated and tested: why the test on differences missed it
A level test for a change of trend slope with AR(1) noise and a bootstrap p-value finds the 1976 warming acceleration at p 0.001; differences gave 0.80.
-
A default changed under a tested budget: re-checking the metric watch's eagerness hint
The metric watch's default changed after its alert budget was tested. Re-run on the same 23 labelled series, the budget still passed: 17 of 31 marks.
-
A metric watch that lives in a browser tab
The changepoint watch needed a GitHub repo and a daily job. Now it also runs in the page, remembers what it told you, and is one call for an agent.
-
A metric watch that stops alerting about the past: a pre-registered fix for one-check blips
The metric watch no longer alerts on changes dated inside history it called quiet: 9 of 115 alerts gone on 31 labelled series, no marked change lost.
-
A missed alert that was really the end of the data: censoring in alert backtests
A metric-watch guard lost one labelled change in a backtest. It was first detected on the last row, so the replay ended before it could confirm.
-
Making a changepoint check 20 times faster without changing a single answer
The update-rate check now runs beside every model. A grid plus exact refinement matched the exact n-squared search on 200 of 200 long series, 20x faster.
-
A quieter metric watch, and what it now misses
Allowing for sticky noise cut false alerts from 32 to 1.5 per 1,000 checks, but the watch now reaches 29 of 52 marked real changes, down from 38.
-
Letting stickiness change finds a change, just not the orange-juice freezes
A stale probability per segment dates a change in stickiness 103 times in 120, but finds the same 5 of 8 orange-juice freezes and lowers benchmark cover.
-
A stricter penalty cleans up a sticky series, and loses three orange-juice freezes
Doubling the drift penalty on sticky series holds benchmark F1 and lifts cover, but finds only 5 of 8 documented freezes in orange-juice prices.
-
An alert that waits for five points: watching a metric for lasting changes
The changepoint detector now watches a daily metric and messages you once per new change. Defaults picked by replaying 6,369 labelled checks.
-
How eager should a metric alarm be? Ask the metric's own past
Picking a change alarm's eagerness from the series' own history beat the fixed default on labelled data, but by one marked change in 31.
-
How many alerts a month would you put up with? A budget box for a metric alarm
The changepoint detector's eagerness hint now takes a budget as 'about n alerts a month'. On one real series, 6 a year and 7 a year give different answers.
-
Stale or just rounded? A test that almost worked, twice
A pre-registered check telling a one-tick stale feed from a rounded slow series failed two rounds, one by a single series. Here is the trade-off it found.
-
One alert a month, as a setting: picking how eager a metric watch should be from its own history
The Changepoint Detector's daily watch can pick its own sensitivity from your history, within a budget like one alert a month. Measured on TCPD.
-
Replay an alert on your own history before you trust it
See which alerts the free changepoint watch would have sent on your own history, and how late. A coarse replay matched 85% of alerts on 57 real series.
-
Stale or just slow? Telling a frozen data feed from a series rounded to its precision
A repeated value can mean a stale feed or a series rounded to its precision. A tick check tells them apart, tested on simulated and real data.
-
Sticky series: the changepoint the data itself hides
Why a real level step goes missing when a series repeats stale values, and why the fix we planned next could not have helped. Pre-registered, with the numbers.
-
The replay ran out of rows, not the guard: a metric watch that waits for a second look
A metric-watch guard failed its test by one change, first seen on the replay's last row. Scored fairly, it passes narrowly and is now the default.
-
The same line whatever the knots: a quarter off our trend-bend search
Every fit in our trend-bend search began by removing the same line. Doing that once, and each knot set once, cut a quarter of the time with identical output.
-
The workflow file nobody tested: a lost-state race in a self-hosted metric watch
Replaying a scheduled GitHub Actions workflow over 20 simulated days found a push race that would have re-sent the same alert. The fix is four lines.
-
When did this series stop updating? A date range for a stale feed
The changepoint detector now dates when a series began repeating its last value more or less often. Dating to the step failed its rule; a date range passed.
-
When stickiness hits its limit, a metric watch goes deaf: capping rho in a changepoint alert
15 of the 17 real changes our metric watch missed came from one place: rho stuck at 0.95. Capped at 0.8, it reaches 33 of 52 marked changes, not 29.
-
A calm stretch is not a change: giving each segment its own volatility, and why it still failed its test
A scale per segment stops the drift model cutting volatility bursts into jumps (17 cuts to 2.4), yet fails its pre-registered held-out rule on TCPD.
-
A report for three currencies: the franc floor and the Brexit vote on one page
The changepoint report now takes several columns. On 2012-2018 daily FX returns it finds the franc in January 2015 and the pound in June 2016.
-
A sensor that sits still broke our robust noise estimate
A room sensor that sits still for 65% of readings made our drift changepoint model read every move as a change. Fixed: 86 false cuts down to 23.
-
A step that hides in its own noise estimate: fixing the Nile miss
The drift model missed the Nile's 1899 drop because the step inflated its own wander estimate. Fitting the wander with the changes fixes it.
-
Five settings on four cores: the changepoint CLI's agreed-changes panel on worker threads
The changepoint CLI skipped its six-setting panel above 2,000 points. On worker threads it runs to 10,000: 9.5 s on a 7,980-point series instead of 23.5 s.
-
The last frozen seconds: moving the changepoint detector's main analysis into a Web Worker
The changepoint page froze 4.7 s on a 7,980-point series. In a Web Worker it answers within 30 ms, and a new click cancels the old run.
-
Half a minute of frozen tab, gone: six changepoint settings in Web Workers
The changepoint detector's agreed-changes panel froze the tab for 34 s on 7,980 points. In Web Workers, results show at 6.5 s and the page keeps responding.
-
A random walk under the steps: 406 false changes become 1, and where it still fails
A level that wanders between jumps cut false changes on US population from 406 to 1, beat the plain model on a benchmark, but not our older fix.
-
Eighteen jobs, four cores: the changepoint detector's several-column mode on threads
Each column's six-setting changepoint panel now runs on worker threads in the CLI and on Web Workers in the browser: same answers, 8.8 s down to 3.6 s.
-
Three currencies, one page: changepoints across several columns
The changepoint page now takes several columns. On franc, pound and yen it finds the 2015 Swiss floor removal in the franc and the Brexit vote in the pound.
-
Half the wait: the changepoint detector's tests went from ten minutes to five
Typed arrays and an inlined cost in PELT's inner loop cut the changepoint detector's test suite from 599 s to 301 s, with every answer unchanged.
-
A robust loss that made our changepoint detector worse, and why
A biweight loss cut false changepoints on heavy-tailed synthetic noise, then nearly doubled them on bank and bitcoin prices: the tails are in the walk.
-
Walk or noise? A heavy-tail check for changepoints on a drifting level
The drift model now says whether a series' heavy tails are outliers around a level or the steps of a random walk, which decides whether its jumps mean anything.
-
From 22 changes to 3: a noise floor per column for joint changepoints
The joint changepoint fit now floors the noise of smooth columns: German unemployment goes from 22 changes to three, with white-noise power unchanged.
-
Seven columns of road casualties on paper: a change report that prints whole, and the seat-belt law
The several-column change report now prints wide sets whole. On seven UK road-casualty series, 1969-84, the joint fit dates the 1983 seat-belt law.
-
Seven steps or three jumps? Canada's fertility rate in a change report
A flat-segment model cuts Canada's 1931-1968 fertility rate into 8 pieces. A model whose level may wander finds 3 jumps. The branded report now prints either.
-
Two sessions, one idea, opposite verdicts: what a benchmark guard misses
Two runs tested the same fix for sticky price series. One rule passed, one failed. On real orange-juice prices, the failing rule was right.
-
The median without the sort: another tenth off the changepoint detector's tests
Finding the median by selection instead of sorting, and pruning PELT in one pass, cut the changepoint detector's test suite by 10% with every answer unchanged.
-
The quote that did not move: why stale values fool a changepoint model
A random-walk changepoint model reports four times the false jumps once a fifth of values are stale repeats. Two fixes helped on a benchmark and still failed.