RDP 9606: The Information Content of Financial Aggregates in Australia 6. Discussion of Results

We interpret our evidence as indicating that there are no large and obvious correlations between financial aggregates and the variables of interest that can be exploited by policymakers in forecasts using simple VARs. Across the numerous systems we examine, the in-sample and out-of-sample tests do not provide consistent support for the idea that growth rates in financial aggregates contain significant information for explaining subsequent fluctuations in output growth and inflation.

There are isolated instances where certain aggregates contain information in an in-sample setting; however, in no case do we find that any single aggregate bears significant explanatory power across all of the in-sample tests. One example of this finding is the variance decomposition results for the broader aggregates. For the full sample, the placement of the financial aggregate in the causal ordering is crucial to the findings of significance. Specifically, when the aggregate follows the policy targets in the causal ordering, the importance of the aggregates for explaining inflation disappears.

The out-of-sample forecast results indicate that none of the aggregates appear to improve the prediction of real output growth in a real-time setting. On the other hand, the out-of-sample results suggest that some of the financial aggregates may improve the prediction of inflation. The RMSE ratio statistics indicate that models containing either broad money, credit, or currency improve the forecasting of inflation in the two and three-variable systems (and also for the four-variable systems containing credit and currency).^[22]

We suspect that the relationship between inflation and the growth in the financial aggregates has become stronger in the latter part of the sample. This apparent correlation appears to be driving the improvement in the forecasts of inflation in the models where financial aggregates are included. However, figures of the forecasts of inflation 4 periods and 8 periods out of sample (for all aggregates except M3) show obvious improvement in the forecast from the VAR with the aggregate only at the end of the sample. The lack of degrees of freedom prevents us from exploring the out-of-sample forecast performance of these models using only the data from the latter period. The key question is whether these correlations indicate the emergence of a more stable and meaningful relationship between financial aggregates and inflation, or are characteristic only of a particular episode.

Further research is necessary to explore this issue. Aside from waiting for more data, one way to proceed in further examining the usefulness of the aggregates might be to examine forecasting models that employ mixed frequency intervals in order to test whether financial data can improve real-time forecasts of inflation.^[23] Data for real GDP and the CPI are published on a quarterly basis, whereas monetary data are published on a monthly frequency, and released prior to the publication of output and inflation measures. This may give these variables information value that is not captured in a quarterly VAR.

Footnotes

When the exchange rate is included in the system, then models with the financial aggregate actually perform worse that the restricted VAR that excludes the aggregate. Because of the poor out-of-sample forecast results for systems that include the exchange rate (those with and without a financial aggregate), we are hesitant to place much importance on results from these systems. We attribute these results to the random nature of exchange rate changes and the inability of the unrestricted VAR to forecast it adequately. [22]

The availability of financial aggregate figures on a monthly basis may allow the use of a state-space filter as used by Zadrozny (1990) to use higher frequency data to forecast lower frequency variables. This issue is left for future research. [23]