By Andrew Feeney-Seale, Amy Lawford, Becky Sanders and Dominic Tighe
This year’s analysis shows that revisions can be considered as being immaterial for most series tested. This is the same broad conclusion as was reached in the 2016 analysis of 2011-13 data and the 2015 analysis of 2010-12 data.
The results are summarised in the following six tables. Tables 1 and 2 present the analysis of the size of revisions for money and credit and effective rates data, showing the mean revision and mean absolute revision. Tables 3 and 4 present the analysis for bias on the same series. For each series, a T-test is the first test conducted. This is the simplest test for bias, and is a function of the mean and standard deviation of the sample of revisions, and the number of observations.
This test is properly valid only when the sample of revisions is independently and identically distributed; this condition fails if revisions are subject to autocorrelation or to non-constant variance (heteroskedasticity). Alternative tests for bias can also be used: for example, an adjusted t-test can be defined which takes into account any evidence of significant first order autocorrelation in the revisions; and there is the Newey-West test that in this case allows for heteroskedasticity and autocorrelation up to two lagged months. Tables 5 and 6 provide further analysis of the revisions, showing the root mean squared, ratio to mean revised data and ratio of variance for all series.
To view all related tables and Annexes, please download the full article:
For questions relating to this article please contact firstname.lastname@example.org or call +44 (0) 20 3461 5353.