ON THE PERFORMANCE OF SHORT-TERM FORECASTS BY THE JAPANESE ECONOMETRIC MODELS

Abstract

We have investigated the performance of short-term forecasts made by five Japanese econometric models. Four of them are quarterly, and, one is semi-annual. The forecasts are ex post forecasts. In the case of the quarterly models eightperiods forecasts are made as of a given time, and, in the case of the semi-annual model four-periods forecasts are made. Since the comparison among different models involves a fundamental, methodological difficulty, we have tried to ascertain only roughly the general performance of these models by crude methods.
The Janus coefficients show that the forecast errors in the outside-sample period are far larger than those in the within-sample period so much so that the econometric models do substantially worse than the naive model. The naive model used is the autoregressive process fitted to each endogenous variable separately by the least squares. However, the econometric models do better than the naive model in the within-sample period as expected.
In our opinion the most serious drawback in the current econometric models is their instability which the above finding implies. Several checks adopted in Fair [8] are recommended in order to select a more stable model.
We have also found that the severity of recessions is 'underestimated so much so that the recessions do not appear at all in the forecast series obtained from some of the five models.
Scholars abroad are perhaps interested in the performance of the semi-annual model from the standpoint of time-aggregation because models are rarely built on the semi-annual basis in foreign countries. In regard to the performance, however, its forecast differs little from the sum of two quarterly forecasts made by quarterly models.