-
A level that wanders, now on the page: changepoint detection that does not cut a drift into steps
The Changepoint Detector's random-walk level model is now on the page: it runs in a background worker and shows the fitted level, the jumps and a CSV.
-
The changepoint CLI now says the agreed change first, and stops at 2,000 points
The changepoint detector's CLI and change report now lead with the changes six settings agree on. Above 2,000 points the CLI skips it: 32 s vs 6 s.
-
Six changepoint settings, one list: which changes do they all agree on?
changepoint-detector now runs six settings and reports how many agree on each change, with the seeded simulation that measures what that buys.
-
Changepoint detector on a labelled benchmark: two settings near the top, a default that loses to 'no change'
On 31 annotated real series, BOCPD and the consensus panel come within 0.012 F1 of the best published method; the plain level model loses to no change.
-
Checking our changepoint detector against the dates it claims to find
Two textbook changepoint series have documented break dates. We read the source text instead of trusting memory, and the detector's panel agrees on both.
-
Changepoint detection on seven series nobody tuned on
Seven labelled series held out from tuning show which choices of the free changepoint detector hold up: agreed changes do, BOCPD does not.
-
Our changepoint default lost to doing nothing. Here is the warning we shipped instead of a new default
On 31 labelled TCPD series no setting chosen by split half beats reporting no change, but a local noise floor beats our old default on 27 of 28.
-
A panel of six change detectors on three more real series: one hit, one miss
Six change-point settings run on Hurricane Maria deaths, Argentina inflation and coal disasters. A spike is found by all six; a stabilisation is missed.
-
Six changepoint settings agree on the 1973 and 1981 breaks in US inflation
Six changepoint settings on 202 quarters of US inflation all date 1973Q3 and 1981Q4; on Michigan-Huron and Saudi oil the AR(1) settings find nothing.
-
Several columns at once: what a joint changepoint fit buys, and where it over-cuts
On TCPD's three multi-column series, merged per-column agreed changes scored F1 0.521, above published bocpd but below reporting no change.
-
A 25-point line fit is the noise floor that finally stops over-cutting a smooth series
A 25-point line-fit noise floor cuts Australian residents from 4 changes to 1 and keeps a known step on synthetic drift; quadratic fits do worse.
-
Six changepoint settings, four real series: where they agree and where one setting hides a break
Six changepoint settings find the UK seat-belt law and the Nile 1899 drop every time, but a Lake Huron break in only four, as the AR(1) settings find nothing.
-
Can a changepoint detector pick out Hurricane Sandy from 731 days of bike rides?
Testing the changepoint detector's spike labels on Capital Bikeshare data: Sandy ranks first for registered riders, twice the next event; missed for casual.
-
The measles change our detector 'missed' was never there
Four of five TCPD annotators mark no change on weekly measles. Our consensus cut at the 1961 peak; the two fixes do not work as defaults.
-
Dating the UK seat-belt law four ways from monthly casualty counts
Four Gaussian changepoint setups run on 192 months of UK driver casualties all bracket the February 1983 seat-belt law; one dates it to the month.
-
A noise floor that was scale-free on synthetic drift stops being scale-free on real series
A second-difference noise floor ignored straight drift in simulation. On US GDP, Australian population and Mauna Loa CO2 it still over-cuts.
-
Two cuts that date the 2008 crash on S&P 500 monthly returns, and one setting that finds nothing
On 122 monthly S&P 500 returns, the plain mean model dates September 2008 and March 2009; the bootstrap setting finds no change at all.
-
A closure is a fraction, not a count: spike labels on the log scale
Running the spike label on log values finds Hurricane Sandy's bike-share closure for casual riders, where the plain level model found nothing.
-
A bad price is not a regime change: a spike label for the changepoint detector
A mean-shift model reports one bad data point as three level changes. A labels-only --spikes flag now says it is one outlier, tested on gold prices.
-
A curved trend for the changepoint detector, and what it did not fix
--trend quad fits a slope and a curvature under the steps. On Mauna Loa CO2 it cuts false steps from 60 to 46, not to zero.
-
AR(1) correction or a noise floor: which keeps a changepoint tool honest on a drifting series?
On 60 synthetic drifting series with one known step, AR(1) correction removes false cuts but misses the step in a third to two thirds of runs.
-
When the autocorrelation fix erases a real step: Canada's baby boom
On Canada's fertility rate, every changepoint model that may cut dates the 1946 baby boom, but the autocorrelation correction finds no change at all.
-
One tax, two series: a changepoint detector finds the GST in café sales but loses it in cement
Six changepoint models run on 426 months of Australian café turnover all date a change two months after the 2000 GST. The same models lost the event in cement.
-
A changepoint detector cuts a smooth CO2 curve 126 times
Run on 68 years of Mauna Loa CO2, every setting of our changepoint detector invents between 18 and 126 steps in a curve that has none.
-
Two crashes in 33 years of Dow Jones returns: what a changepoint detector finds
On 8,609 daily DJIA returns, the meanvar model dates the 2008 Lehman day exactly and brackets Black Monday 1987. The plain model cuts 27 times.
-
Can a changepoint detector find the 1945 step in global temperature?
On 145 years of global temperature, three changepoint models cut at 1946, the AR(1) variant misses it, and the slope model cannot place the 1976 acceleration.
-
Can a changepoint detector find a hurricane in bikeshare counts?
Count and log models date Hurricane Sandy's two-day bikeshare outage to the exact day; a plain Gaussian model for casual riders is a day off.
-
Which fix for over-cut drifting series finds the step: a noise floor or joint AR(1)?
On 60 seeded drifting series per cell, joint AR(1) finds a known step 60 of 60 times at phi 0.8; plain AR(1) finds 27, a floor leaves 3 false cuts.
-
Does giving a changepoint detector the cycle length stop it cutting the lynx cycle into steps?
On 114 years of Canadian lynx trappings the default finds 23 false steps; setting the period leaves 10 to 24.
-
Does a changepoint detector find the Fed's 1979 to 1982 volatility regime?
On 530 monthly changes of the US 1-month T-bill yield, three variants of the open-source detector all cut at October 1979 and within a month of October 1982.
-
The 1976 warming acceleration: our break test finds the year but cannot confirm it
On 145 years of global temperature, the best slope split is 1976, but the test statistic is 1.8 against a bar of 8.4 (p 0.80).
-
Two ways to stop a changepoint detector cutting a drifting series, compared
On 60 synthetic drifting series with one known step, a noise floor of 1.5 finds the step 59 to 60 times in 60; the AR(1) route finds it 36 to 42 times.
-
What a noise floor buys a changepoint detector on a drifting series
On 60 synthetic drifting series with one known step, a noise floor of 1.5 cuts false changes from 3.6 to 0.3 but loses steps near the size of the drift.
-
A documented step in Australian cement, found only by the models that over-cut
Australia's 2000 GST shows up in quarterly cement output as a fall the changepoint tool dates at 2000Q3 or Q4, but only among 20 or more other cuts.
-
A changepoint tool dates the Crimean War's sanitary fall to the month
On Nightingale's 24 monthly disease death rates, PELT places the fall at March 1855, the month the Sanitary Commission arrived.
-
A real break that only just clears the bar: Australian gas, 1969
The slope-break test finds the 1969 takeoff of Australian gas production in annual totals: W 12.23 against a bar of 12.08, p = 0.048.
-
A changepoint detector on 41 years of Australian travel: Bali and SARS are not found
Monthly Australian resident departures 1976-2017: no seasonal mean model cuts at the Bali bombing or SARS, and the trend model's nearest cut is a month early.
-
Three real shocks in Japanese visitor numbers, and which change-point models see them
Quarterly arrivals to Australia from Japan contain the Asian crisis, SARS and the 2008 crash. No single model in the detector dates all three.
-
The Great Moderation, found again in a second series
The volatility model dates the 1984 fall in US growth volatility on monthly industrial production; one series, a replication, not a calibration.
-
How borderline is a borderline break? Twenty seeds on Australian gas
A real documented takeoff clears a slope-break bar by 1% with one seed. Twenty seeds say the p-value is about 0.045.
-
Two slope bends where US homicide turned, found by a changepoint tool
On 55 years of the US female murder rate, a step-based changepoint model cuts 10 times; a slope-bend model dates two turns, 1980 and 1993.
-
Telling a ramp from noise that sticks: what --ar1 keeps and loses
Synthetic ramps in autocorrelated noise: plain changepoint runs cut 12 to 15 times a series, --ar1 cuts twice, keeping ramps of 24+ points, losing short ones.
-
A 95% range that holds the truth 80% of the time
Replicating the steep-growth cell of the changepoint detector's clumpy-count range on four fresh seeds: 108 of 138 found bends held, 78%.
-
Can a changepoint detector find the Sydney Olympics in visitor numbers?
Quarterly US arrivals to Australia, run through four detector models: two date the Olympics quarter exactly, two lose it.
-
The Marathon Step the Data Does Not Show: a documented change that changepoint detection should not find
We tried the Boston Marathon's 1924 course lengthening as a test for the changepoint detector. The winning times show no step, so no model finds one.
-
The detector dates the 2007 US electricity stall, then declines to claim it
On 40 years of US electricity generation, a monthly model over-cuts 38 times and the slope test finds 2007 but misses its own bar.
-
Australian beer production: the log scale dates the 1976 bend 15 years tighter
On 218 quarters of Australian beer production, the log series puts a slope change at 1976Q2 within three years; the raw series spans fifteen.
-
One ramp, not three steps: grouping a changepoint detector's cuts
A changepoint detector cuts a smooth rise into steps. The changepoint-detector reports adjacent same-direction cuts as one ramp, a grouping rule, not a test.
-
Ramp or two steps? Letting the data pick, on simulated series
Replacing a fixed gap rule with a BIC comparison of a straight ramp against steps finds long ramps in 33 to 40 of 40 simulated series instead of 0 to 8.
-
Ramp or two steps? A gap rule that cannot tell, measured
Grouping nearby same-direction changepoints into one ramp finds short ramps, misses long ones, and cannot separate a ramp from two close steps.