A series that repeats its last value can mean two different things. Either nobody refreshed the figure (a stale feed, a price carried forward), or the thing really moved less than the precision it was recorded to (a slow index stored to 0.1). The Changepoint Detector's update-rate check already tells most cases apart by how big the moves are: if at least half the moves in the stickier stretch are a single tick, it says the repeats may only be rounding. Its README admitted the hole. A stale feed whose real moves are themselves about one tick passes that test too, so it gets called rounding.
The idea
Look at the move that ends a run of repeats. A rounded slow walk leaves a run by crossing the next rounding boundary, so that move is still about one tick however long the run was. A stale figure catches up on every step it skipped, so after k repeats its move has about k + 1 times the variance. The statistic is the mean squared move ending a run of repeats, divided by the mean squared move that follows an ordinary move. It comes out near 1 for rounding and about 2 to 4 for a one-tick stale feed.
Two rounds, both written down first
Each round's pass rule was committed before its confirmation seed ran. On 300-step simulated series with a change halfway through:
- Round one (runs of 2 or more, ratio 1.4): it caught 90 to 96% of the one-tick stale feeds. But a rounded walk of 0.6 ticks a step was called stale in 6.0% of series, and the bar was 5%. Fail.
- Round two (setting picked on a fresh calibration seed with a 2.5% ceiling for false calls; runs of 3 or more, ratio 1.8): false calls fell to 1.5%, and two stale cells passed at 85%. The third caught 79.5%, 116 of 146, against a bar of 80%. Fail, by one series.
There is no third round. Tuning again on the same cells until something passes is exactly the forking path these rules exist to stop. So the check is not offered: every sentence the tool writes is unchanged, and each change only carries the ratio as a plain number.
What it did show
The two rounds bracket a real trade-off. With about 300 steps a stretch holds too few catch-ups for this statistic to give both 90% power and a 5% false-call rate. On real data it behaves as the reasoning says: the producer price index in the frozen orange juice file, a rounded series, has a ratio of 0.48. The move that ends a run there is smaller than a typical move. The juice price itself (already called stale on move size) comes in at 1.73: catch-ups larger than ordinary moves, but less than a carried-forward random walk predicts. Why is an open question.
Next: a power curve in series length, which would show where one setting could pass both bars, or a proper likelihood comparing a rounded walk with a rounded walk carried forward. The tables, the code and both rules are in the README under "One-tick stale feeds".
The Changepoint Detector is free and open source (MIT). Its output describes data; it is not financial, scientific or professional advice.