This post was supposed to be released at the beginning of next week. However, given this morning figure from Germany (new orders, sa) we anticipated it. The reason is that many series have lost or reduced their pre-Covid seasonal pattern. Therefore, the end-of-sample distortions are way more pronounced than in the past. We discuss (via Q&As) this issue using core HICP as an example, and we offer a simple suggestion. A similar reasoning applies to other series in other sectors. The conclusion is the following: there is no easy fix, the data are prone to larger-than-usual revisions, and each series should be treated as a special case that requires an ad-hoc treatment. Our recommendation in this environment is to take little or no signal from the last observation of MoM sa, unless one is really sure about how that number is constructed.
(Stop here if not interested in seasonal adjustment issues)
Different seasonal adjustment methods give different answers. Why? And which one is the best one?
Each seasonal adjustment method is a different model with different assumptions and gives different answers. Especially in the current environment, there is no “easy” fix. The performance of a filter strictly depends on the dimension to evaluate. We reiterate our suggestion: one should not rely on “built-in” procedures if (s)he is not familiar with the econometrics behind it. Let us make an example: so many people assume that the Hodrick-Prescott (HP) filter is a good technique. Unfortunately, this is false and the use of the HP filter is never recommended. There is a reason why James Hamilton wrote a paper titled “Why You Should Never Use the Hodrick-Prescott Filter”, even if many seem unaware of it. Also, one should doubt of seasonally adjusted data released by statistical agencies right now because many series have lost the pre-Covid seasonal pattern and the agencies have not revised their methods yet. This implies that in some cases the agencies are (most likely) introducing noise/distortions in seasonally adjusted data. (And yes, there is no certainty that using Census X-13, which is generally considered “the standard” results in non-distorted series)
Is the UnderlyingInflation series of core HICP “better” than the ECB seasonally-adjusted series?
The (MoM) seasonally adjusted number we estimate of core HICP provides a better real-time estimate of the “true” signal. As discussed last December (see here) our own seasonally adjusted series is unbiased, less volatile than the alternatives (ECB, etc..), and does not suffer from residual seasonality. Figure 1 shows a comparison of the MoM seasonally adjusted of core HICP across filters. As discussed last month (see here) we did not take any signal from the softness of the MoM estimated in March by the ECB series (as well as other methods). Once again, it was the right choice, given that the ECB series bounced back in April (for the record, and unsurprisingly so, the cumulative basis points Feb-Apr of the ECB series is 151 vs 149 of our own series. Translated: on a MoM, our suggestion is to largely discount any large movement of the ECB series: the fact that the MoM of the ECB series accelerated in April does not mean that inflation is, in fact, accelerating. We are most likely going sideways-ish, as shown in Figure 1).
Figure 1. MoM of seasonally adjusted core HICP using different filters
Another way of making the same point is to look at Figure 2 which plots the level of HICP core goods (“Non-Energy Industrial Goods” – NEIG). The blue line shows the level seasonally adjusted by us, while the gray line shows the level of the ECB series. It is pretty clear that the gray line has a tendency to fluctuate around the blue line, that is the ECB series contains some residual seasonality which is eliminated in our series.
Figure 2. Level of seasonally adjusted HICP NEIG
How does the ECB staff come up with the SA series? Does the Governing Council rely on that?
To our knowledge, the exact procedure (that is, the exact set of assumptions) behind the ECB seasonally adjusted series is not public. Some information can be found here, here, and here. But again, the devil is in the details and “X-12” or “X-13” says little because they indicate classes of models, not a well-defined single procedure. Small differences in assumptions produce large differences in the results. Also, in our experience the ECB staff and the Governing Council members used to rely almost exclusively on the YoY of the series (from which, no need to carefully seasonally adjust it). Things are changing, it seems. But overall, we continue to think that a large number of people among the ECB staff and Governing Council members continue to erroneously rely on the YoY.
Will the filters revise again recent history next month? And by how much?
Absolutely yes, but it is very hard to say how much and in what direction. Any filtering procedure revises history each time a new observation is added. Unfortunately, in the current environment the revisions are larger than usual because the end-of-sample distortion is larger than usual (see our note about Spanish CPI for a discussion about missing seasonality). Next month, the filter(s) will revise recent history again. Unfortunately, it is virtually impossible to know the direction of the revision without the report, although risks are probably skewed to the upside. Our suggestion: if one would like to try to anticipate what will happen, one can simulate the NSA level in the next 3-4 months and re-run the filters, as we did for the Spanish CPI. Our research suggests that the MoM in the near-term should remain solid (0.4%-ish). But again, there is no “easy” answer and there is no certainty, of course.
Any good news?
Yes, look at the 3m/3m. There is good news: while the MoM of the seasonally adjusted series typically send conflicting signals across filters, smoothing the level of the series helps a lot. Indeed, the 3m/3m changes across methods generally results in smaller differences, as shown in Figure 3. Therefore, for any practical purposes on a daily basis we recommend to construct the 3m/3m (better than the 3m) of a seasonally adjusted series and monitor that. Of course, a better approach would be to treat each series as a unique case and consider all its peculiarities (i.e. do the holidays affect that series as they do with the other? Does the number of weekend days play a role? Does the weather play a role? How do you cope with missing seasonality? How do you correct for residual seasonality/calendarity? Etc..) but admittedly that requires a big investment.
Figure 3. 3m/3m ar of seasonally adjusted core HICP using different filters
What is your forecast for the next month?
Keep it easy, it works. In the last several months we have seen several sell-side analysts forecasting a sharp deceleration of the MoM sa (in some cases, as low as 0.2% or even 0.1%). We have respectfully disagreed, and it did pay off. The reason is simple: the MoM forecast must square with the distribution of price changes and the macro models. Right now, the distribution is centered around 5%-6% ar. Therefore, is it really surprising that prices have been growing at about 6% ar in the last few months? To us, it is not.
(Again, a similar reasoning applies to many other series, including the ones released this morning for Germany)