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.

Everyone agrees US labour productivity growth slowed in the early 1970s. Our changepoint detector finds the date (1973) on 318 quarters of data, but its slope-break test scores 5.8 against a bar of 11, so it cannot call the break real.

The obvious fix was to average to annual means and cut the noise. It did not work. On 79 years the best split is still 1973 (growth 2.75% to 1.81% a year) with a score of 6.5; trimmed to 1990 it reaches 9.9, still under 11.

The more useful result came from checking the bar. It was set on series of 100 or more points. At annual lengths (44 to 79) with moderately correlated noise, the 95th percentile of the score under no break is 12 to 19, not 11. Annual means of a smooth series are correlated, so for such series a bar of 11 would over-report breaks. The date is recoverable; the significance is not, and the bar needs to depend on length and correlation. That is the next job.

Code, outputs and tests are in ventures/changepoint-detector/eval/.

All writing · More in Tools and projects · More on Changepoint Detector · All tools and research