While it is still early in the round, given yesterday’s CPI numbers we have updated our “main” model (as a reminder, the “main” model is an augmented Phillips curve model as described in Detmeister et al. (2014) and assumes a flat pi* both in history and in the forecast). As usual, our forecast includes a point forecast, as well as two types of confidence bands: A. confidence bands calculated using 100,000 draws from estimated parameters’ distributions, and B. confidence bands on the model’s historical forecast errors (quasi out-of-sample exercise).
The sample includes our 2021:Q4 nowcast (+4.5% at annual rate) and the forecast extends to 2024:Q4. Compared to the last update, the starting point of the forecast is a bit higher and we have a slightly tighter labor market (please note that the starting point is only a “bit” higher because the model was already forecasting a relatively strong Q4 number – +2.9% in the last forecast).
Results
The inclusion of 2021:Q4 in the sample has triggered an additional upward revision to the already elevated forecast of core PCE price inflation. The model is currently forecasting elevated inflation not only in 2022 but throughout the entire forecast horizon (with upward risks around the point forecast).
Comparison with previous forecasts
Given the current model forecast, it is instructive to compare it to the forecast stopping the estimation in 2021:Q1. The output of this exercise is reported in the Appendix. In a nut shell, stopping the estimation before the recent runup results in a forecast that shows the typical feature of an augmented Phillips curve model: (i) the model’s forecast is negatively serially correlated, (ii) the forecast converges back around the anchor in a relatively limited number of quarters, and (iii) the confidence bands suggest moderate uncertainty. On the other hand, the current forecast does not display any negative serial correlation, the forecast remains well above the anchor throughout the entire forecast horizon, and the confidence bands suggest very high uncertainty. What explains the different properties of the forecast? What happened in the model is that the recent quarters are large enough (and positively serially correlated) that the estimated autoregressive coefficients of the model are now all positive (with the exception of one lag which is still negative but close to zero). Please note that this happened despite the fact that the model is estimated over the 1996 to present sample. For this reason, the forecast remains elevated and it takes a lot of time to converge back around 2 percent. This also implies that (most likely) going forward the model will continue to produce elevated forecasts, unless we will finally get a sequence of low (or negative) quarters.
A word of caution about the model forecast
Our “main” model is a workhorse augmented Phillips curve model, extremely popular in central banks. Nevertheless, at any given point in time the information set of the econometrician is larger than the model. In this moment the information set of the econometrician is based on the narrative of the incoming data, especially about the level of durable goods. Put it differently, there is a reason to believe that the model might be wrong going forward but only if one think that the level of (durable) goods will drop significantly in the upcoming quarters.
Implications for the Fed Board staff
We employ our “main” model as a way to assess the risks around the Fed staff forecast. Under normal circumstances (again, see Figures in the Appendix), there is only a small difference between the model’s forecast and the staff forecast (which, as known, remains fully judgmental). The Fed Board staff forecast is significantly lower in 2022 (and beyond) than the “main” model despite the fundamental assumptions -such as the level of underlying inflation- are the same. As previously discussed, the staff is working under the assumption that its own information set is larger than the model. Under the staff view, most of the recent strength will prove transitory, as the level of durable goods will reverse in the upcoming quarters and services inflation will moderate. The staff might prove to be right; however, in this moment we think the staff forecast does not balances the risks. Indeed, all our models (Phillips curve – thick modelling – trend models) suggest that the risks around the staff forecast lie to the upside.
Figures
Inflation forecast (YoY of core PCE, %) from "main" model – latest run (2021:Q4 in sample)
A. C.I. calculated from parameters distributions
Note: the confidence intervals (C.I.) are estimated using 100,000 draws from the estimated parameters distributions to simulate the path of core PCE price inflation going forward.
First quarter of forecast: 2022:Q1.
B. C.I. calculated using quasi-out-of-sample methods
Note: the confidence intervals (C.I.) are estimated using quasi-out-of-sample methods (estimate the model over a sub-sample, forecast, and calculate the root mean squared forecast errors).
First quarter of forecast: 2022:Q1.
Appendix
Inflation forecast (YoY of core PCE, %) from "main" model – stopping the estimation in 2021:Q1.
A. C.I. calculated from parameters distributions
Note: the confidence intervals (C.I.) are estimated using 100,000 draws from the estimated parameters distributions to simulate the path of core PCE price inflation going forward.
First quarter of forecast: 2021:Q2.
B. C.I. calculated using quasi-out-of-sample methods
Note: the confidence intervals (C.I.) are estimated using quasi-out-of-sample methods (estimate the model over a sub-sample, forecast, and calculate the root mean squared forecast errors).
First quarter of forecast: 2022:Q1.