The December MoM of core HICP is estimated between 26bps and 49bps according to seasonal adjustment procedures. Because the consequences for monetary policy can be different, which one is the “true” signal? In this note, we discuss why a researcher should be skeptical of “built-in” seasonal adjustment procedures, we argue that the December reading was very strong, and why we can expect another very strong reading in January.
Details
Census X-13 indicates a class of models with “N” options, not a well-defined single method. Two researchers using “Census X-13” get different results depending on the chosen options, the sample size or the treatment of holidays. This explains why, for instance, built-in seasonal adjustment methods in data providers (i.e. Haver vs Macrobond) result in different seasonally adjusted series. The reader interested in the details can refer to Census (2017), BLS Tiller and Chow (2007) or BLS Tiller et al. (2005) for a discussion on X11 vs SEATS (and why SEATS is generally preferable). The bottom line is that unless a researcher understands the details and agrees with the chosen procedure, one should be skeptical of “built-in” seasonal adjustment methods or at least one should compare across options. Taking Macrobond as a reference, in the next section we show how significant the differences and the issues across seasonal adjustment procedures can be, including in the series published by the ECB.
We provide a seasonally adjusted series with superior properties. Figure 1 shows the MoM of core HICP comparing 4 seasonal adjustment methods: X-11 as embedded in Macrobond (“MB X-11”), SEATS as embedded in Macrobond (“MB SEATS”), the seasonally adjusted series released by the ECB (“ECB” which employs X-12), and our own (“Underlying Inflation” which relies on a double X-13 method). Table 1 reports three statistics of the four series: the mean, the standard deviation, and the share of months in which the MoM is negative. The bottom line is that both X-11 and SEATS result in a series that is quite volatile and negative (that is, in deflation) in 11.7% and 8.0% of the months, respectively. On the other hand, the ECB series is negative “only” in 5.7% of the months. Finally, our in-house seasonally adjusted series is unbiased, the least volatile, and almost never negative (which is what one should expect, given downwardly rigid prices).
(For the record, looking at longer horizons such as the 3m/3m instead of the MoM does not solve the problem because the gap across series can be as large as 1 percent (that is, one measure running at 1% a.r. and another at 2% a.r. – chart available upon request)).
Figure 1. MoM of seasonally adjusted core HICP
Note: the figure shows the MoM of core HICP estimated using 4 different seasonal adjustment methods. “MB” stands for “Macrobond”, “ECB” refers to the series published by the ECB (here).
Table 1. MoM of seasonally adjusted core HICP, statistics
All series suffer from “residual seasonality”, except for our own. Figure 2 shows the mean and confidence intervals of the MoM of core HICP across months of the year and seasonal adjustment methods. Ideally, a seasonal adjustment procedure removes all seasonality and results in a uniform distribution across months. Instead, X-11, SEATS and the ECB produce series with “residual seasonality”. For instance, a regression of the ECB MoM on a set of monthly dummies results in 5 coefficients statistically significant, with January above all other months of the year (as seen in Figure 2, bottom left panel). On the other hand, our own series shows no residual seasonality and a distribution across months close to ideal (same mean and standard deviation in each month).
Figure 2. Mean and Confidence Intervals of MoM of seasonally adjusted core HICP
Note: the figure shows the average and confidence intervals of the MoM of core HICP across months of the year and seasonal adjustment methods. Ideally, a seasonal adjustment procedure results in a uniform mean and standard deviation across the 12 months.
Expect another very strong reading (0.4% MoM, sa) in January. In December, the MoM of core HICP is estimated between 0.4% and 0.5% across the 4 methods discussed in this note, above the 0.3% suggested by some other built-in procedures in data providers. All told, there are two conclusions. First, our experience and estimates suggest to be skeptical of any seasonally adjusted HICP figure, unless you run your own model and control the assumptions. In this dimension, every model has its own limitations; therefore, comparing estimates across models (especially our own series, given its properties) is always a good idea. Second, if you rely on the ECB series you should expect another very strong reading in January because of residual seasonality next month. This is also confirmed by the fact that in December the MoM of our own series was a bit higher than the alternatives (49bps vs 39bps using MB X-11, 48bps using MB SEATS, and 42bps looking at the ECB series). Given the persistency of the series, the next print could be another solid 0.4% MoM (sa).