A variance-change model has a trap: a short run of similar numbers has a tiny sample variance, so the model reads it as its own calm regime. On 202 quarters of US GDP growth, changepoint-detector's meanvar model with a minimum segment of 5 reported a five-quarter calm patch in 1983-84 and dated the Great Moderation to a range that missed the documented date.
That was one real series. This run checked the minimum segment on simulated series with a known truth (eval/meanvar-min-segment.mjs, seeded, 100 per cell):
- With no break, false alarms drop from 3% to 2% (n = 60) with a minimum of 10.
- With one sd break at the midpoint, the share of runs that find exactly one change rises from 62-67% to 69-72% at n = 60, and from about 86% to 90% at n = 100. Dating error and the 95% range hit rate (94-98%) stay put.
- The price: an 8-point burst of sd 4 inside sd 1 is found by 73% of runs at a minimum of 10, against 77% at 5. At 12 it falls to 30%.
So meanvar now defaults to a minimum segment of 10, which also gives the correct range on the GDP series. Pass --min-segment 5 for short regimes. The simulation is Gaussian only; heavy tails and autocorrelated variance are still untested.