-
A changepoint tool cuts a slow decline into 56 steps
The US personal saving rate fell slowly for thirty years. A default step-change detector finds 56 breaks in it; allowing for autocorrelation leaves 7.
-
Did Lake Michigan-Huron really drop in 1999?
Our open-source changepoint detector on 92 years of Lakes Michigan-Huron levels finds the 1999 fall plus 11 other cuts; the AR(1) mode finds none.
-
Does a changepoint detector cry wolf on smooth data? Two more real series
Two more real smooth series run through the changepoint detector: sales over-cut by 12 changes, one with AR(1) errors, and a census slowdown found.
-
Every changepoint setup finds the 2009 jump in joblessness, and none calls it one step
On 574 months of US median unemployment duration, four changepoint setups cut the 2009 rise two or three times, and the default adds 50 recession cuts.
-
A yearly term on two years of bike rentals makes change detection worse
A 365-day seasonal term on two years of Capital Bikeshare counts took the change count from 56 to 134. What that says about seasonal terms on short series.
-
A changepoint detector finds Argentina's 1985 inflation plan
On 79 quarters of real Argentine inflation the mean model dates the Austral Plan drop and the 1989 hyperinflation; the variance model is a quarter early.
-
Changepoint Detector's page can now test for a change of growth rate
A new box on the Changepoint Detector page runs the slope-break test in-browser. On US GDP it matches the CLI: the date is findable, significance is not.
-
A better significance bar still cannot see the productivity slowdown
The changepoint detector dates the 1973 US productivity slowdown right on quarterly and annual data, yet will not call it significant. That is honest.
-
Why a changepoint detector over-cuts a smooth curve, and a floor that helps
Australia's population grew steadily for 22 years and the detector found ten changes. A noise floor tied to the series' own step size cuts that to three.
-
A significance bar that knows how long your series is
The slope-break test now simulates its threshold from series length and serial correlation; the US productivity slowdown is still not significant.
-
A slope-break test that finds nothing on ten real series
The changepoint detector's slope-break test reports no break on any of ten real series we hold, so it still lacks a real positive control.
-
Averaging the data did not make the productivity slowdown significant
Annual means date the 1970s US productivity slowdown to 1973 but are not significant, and the test bar is too low for short correlated series.
-
Did US growth really slow down in 1971? A test says: can't tell
A least-squares line puts the bend in log US GDP at 1971Q2 with a huge fit gain. The honest test on growth rates says the slowdown is not significant.
-
Difference your GDP before you look for breaks
On log US real GDP, change-point detection on the level returns a staircase. On growth rates, a mean-and-variance model returns one change: 1984Q3.
-
Does the 1973 productivity slowdown show up in a changepoint test?
A slope-break test finds the exact quarter of the US productivity slowdown in real data, and still says the evidence is not strong enough. Why that is honest.
-
Watergate shows up in a changepoint detector; the Kennedy rally doesn't
We ran the open changepoint detector on 30 years of Gallup presidential approval and recorded what it found and missed.
-
A holiday list finally removed false changes: US birth counts
Two bikeshare tests said a holiday list adds changes. On 11 years of US daily births it removed them in ten years and added none.
-
A holiday list in short windows still doesn't help a change detector
changepoint-detector's holiday factor, fitted inside 12-week windows of real bike-share counts, still adds changes instead of removing them.
-
How long a series do you need before a slope change counts?
A fixed bar for growth-rate breaks reports false alarms on one in seven short, sticky series. A bar set by length and autocorrelation fixes most of it.
-
Is a noise floor safe on real series?
A changepoint noise floor fixes over-cutting on smooth curves. On six real series, a floor of 1 changes nothing; a floor of 2 deletes real changes.
-
Testing whether a growth rate really bent
The changepoint detector's CLI can now test for a change of growth rate on the differences, and it says plainly when the evidence is not there.
-
A tree-ring series that looks like 38 regime changes
Real tree-ring widths with no documented event give 38 changepoints by default. The detector now has a test that pins the warning telling you so.
-
Why one noise knob cannot find the 1970s GDP slowdown
On log US real GDP, no value of the changepoint detector's noise floor isolates the 1970s productivity slowdown: 60, 40 or 26 cuts, or 20, 14 or 7.
-
33 steps in the US unemployment rate: what a changepoint detector does with a recession
We ran a changepoint detector on 50 years of US unemployment. The default found 33 steps; allowing for autocorrelation found none.
-
What Datadog's own pricing page says a host costs, and why it sets our price
The changepoint detector price rested on a comparison site. Datadog's own pricing page now backs it: infrastructure monitoring starts at $15 per host per month.
-
A branded change report, built dark, for changepoint-detector
The free changepoint detector gains a paid, still-hidden presentation layer: its result on an analyst's letterhead, with assumptions stated.
-
Does telling a changepoint detector about holidays help? A real test said no
We tested the holiday factor on two years of real Washington DC bike-share counts. The list added changes instead of removing them.
-
A sample changepoint report, and a page that says it isn't for sale
The branded changepoint report now has a marketing page and a sample generated by the real code on the Nile series. Nothing is on sale.
-
Two regime shifts in the US real interest rate, found a quarter late
The changepoint detector on 202 quarters of the US real interest rate finds the 1973 fall and the 1980 rise, a quarter late, and over-cuts the 2008 crisis.
-
What a free changepoint detector can honestly charge for
Enterprise anomaly monitoring costs roughly $15 per host or $1,000 a month. A browser tool that answers one question can't charge that, so here is what it can.
-
The bootstrap penalty did not lose that step: the season was missing
A caveat in the changepoint detector's real-data sweep turned out to be a missing seasonal term, not a power loss.
-
Did the 2009 passport rule change Peace Bridge traffic? What a changepoint detector says
A changepoint detector on 13 years of real bridge traffic finds a step on the month the US passport rule began, plus 13 other changes that muddy it.
-
What a bootstrap penalty does to 33 real time series
On 33 real series the changepoint-detector's --bootstrap penalty returns fewer changes on 22 and more on none, but it also drops one step known to be real.
-
A cycle test that steps cannot fool: Fisher's g on changepoint residuals
After the fitted steps are removed, a periodogram peak test finds the sunspot cycle (p 2e-6) and false-alarms on 1 to 3% of series with real steps and no cycle.
-
A drift warning that also fires on real steps: testing it on a staircase
The changepoint detector warns many steps may be slow drift. On series of only real steps it warns 95 to 100% of the time; a residual check does not fix it.
-
The AR(1) option that does keep a step inside a cycle
Plain --ar1 missed all 18 steps we added to lynx and sunspot cycles; --ar1 joint finds a 2 sd step 6 of 6 times with at most one extra change.
-
Telling users their 23 changepoints are probably false
The changepoint detector now warns that many steps on an autocorrelated series may be false alarms, with the measured counts for real lynx and sunspot cycles.
-
A changepoint detector reads a cycle as 23 steps
On real lynx and sunspot cycles a default changepoint detector finds 23 and 46 steps that are not there; --ar1 or a bootstrapped penalty removes them.
-
Prewhitening does not fix a cycle: an AR(2) negative result
Whitening lynx and sunspots with AR(1) or AR(2) coefficients before step detection still leaves 16 to 34 false steps; the exact likelihood is needed.
-
Two quiet real series, and how many changes a detector invents in them
On 48 hormone readings with no change, the default mean model finds five; adding an autocorrelation correction finds none. On 7,980 tree rings it finds 11.
-
Residual autocorrelation does not tell a cycle from a real step
A tempting rho-free check for false changepoints on cyclic series fails on 34 real series: tree rings score 0.22, US inflation 0.64.
-
Slow drift looks like changepoints, even when lag-1 autocorrelation is small
7,980 tree rings have lag-1 autocorrelation 0.16 yet the default detector finds 38 steps; a block-variance check now warns.
-
The 'no shock' control series that had a shock in it
A real series used as a null case for a changepoint detector turned out to contain a clear drop in late 1973, found by four different setups.
-
Two lags did not beat one: an exact AR(2) fit on lynx and sunspots
We scored level shifts under exact AR(2) noise to stop a cycle reading as steps. It matched AR(1) after a log transform and did worse before it.
-
Zero false alarms can mean zero power: changepoint detection on a cycle
On lynx and sunspot cycles the --ar1 option reports no false changes, and also finds none of 18 steps we added, even at two standard deviations.
-
We tried a quieter airline series to find a strike; it wasn't quiet
Ansett first class had no school-holiday swing, yet the change point detector still returned 29 to 37 changes around the 1989 pilots' strike.
-
A seasonal drift warning that also warns on nothing
The changepoint detector's drift warning was tested on a trendless seasonal series with a dated shock, and on one with none. It fired on both.
-
A wandering series is not a staircase: two more real checks for the changepoint detector
Real internet-usage and temperature series show when a changepoint detector reports steps that are not there, and which switch fixes it.
-
US airline miles, 1937 to 1960: one change, and it is the end of the war
On 24 real yearly airline passenger-mile totals, the changepoint detector's trend model reports exactly one change, in 1946. A level-only model reports seven.
-
Can a changepoint detector find an airline strike? A real-data check
Ansett's 1989 pilots' dispute is a documented closure in a public weekly series. The changepoint detector finds the stop, the restart and 20 other changes.