Over the past weekend, I read a column on ECB’s measures on underlying inflation (this is probably something I will return to in a future blog post) on one of my favorite sites, which lead me to a couple of interesting boxes in the 4th issue of the ECB’s Economic Bulletin from this summer. That reminded me of a Liberty Street blog post (FRBNY) introducing the FED’s ‘Underlying Inflation Gauge’ (UIG) that I know we added to the MB database and I somehow forgot to explore. I had to pick up the laptop and look closer…
– I should think that just by eyeing the chart above we can agree that there appears to be a lag structure that is ripe for exploring. First, I just used the correlation analysis to find the best lag-lengths.
It would seem that the ‘Total’-measure not only demonstrates higher average correlation, it also seems to lead core-CPI by a little over a year. Currently the lags are 14 months for total UIG and nine months for ‘Prices-Only’ UIG (the difference between the two is explained in the initial links). Here, I tried different estimation sample-ranges and they suggested some sort of break before and after the great recession. It seems to be more pronounced for the prices-only measure though. It is only in the very latest period that the total measure struggles more noticeably and the lag-structure compresses. This can also be shown using the covariation analysis in the Macrobond application (here I have used the full sample lag structure).
One way of further exploring these measures and inherent idiosyncrasies would be by using the rolling regression. However, given the construction of these measures the use of a rolling regression should add little benefit and the statistical output of the normal regression analysis is much richer. Plus, it’s the only way we get to do dynamic forecasts, which look intriguing in any presentation or analysis.
Before going all haywire about the model’s predictions, do note that the tests indicate strong positive serial correlation (DW-stat) which implies inflated t-statistics; too low standard errors. In practice, this means that while the point estimate may be fair, the uncertainties surrounding it – the risks of model errors – might encompass very different scenarios than what any point estimate indicate. That said, using year-over-year differences poses somewhat (small understatement) of a challenge to the DW-statistics (and we can also reject the unit root test).
Now, I am far too old and jaded to get fired-up about the results of these type of exercises, but at least it provides another angle on why the FED by and large seems set to hike rates once a quarter for the foreseeable future (in this case a little over a year – at least).
 I do look forward to the publication of Bobeica et al: “PCCI – a data rich measure of underlying inflation for the euro area” in ECB Statistics Paper Series.
 Just so that you don’t get your hopes up, adjusting/controlling for serial correlation in year-over-year data is beyond the scope of the MB-applications statistical functionality. No need to send support tickets on this. – If this is something you need to do, better then to test data in other transformations (or even, heaven forbid, other software).