RDP 2001-09: What do Sentiment Surveys Measure? 6. Conclusion

To summarise, we find that when readily available economic information is appropriately ‘filtered’ from sentiment indices, these indices mostly fail even rudimentary Granger causality tests of predictive ability. The residuals that we would expect to be the most forward-looking (i.e., those extracted from regressions of forward-looking indices) often turn out to be no more useful for predicting the future than those from backward-looking indices. Thus, there is reason to suspect that respondents' forecasts offer little more information about the future path of the economy than a weighted average of lagged economic variables. While sentiment indices appear to explain some of the variation in the growth of GDP, employment and household expenditure, there is no a priori basis for presuming that the indices' implicit weighting of economic information is consistently better than other possible weighting schemes. However, if asked to identify the most useful indices among those analysed, we would choose the NAB Actual Business Conditions and Expected Employment 3-month Outlook indices, the Roy Morgan average index of consumer sentiment, and the second component index (expected personal financial conditions) from both consumer confidence surveys.[23] These indicators have some predictive ability for employment growth and (in the case of Index 2 from the consumer confidence surveys) the probability of a recession.

We conclude that while sentiment indicators may provide a rough summary of available economic information, it would be risky to claim much more on their behalf. An investigation of the predictive power of component indices corresponding to the individual survey questions yields limited further insight. Though we document two main exceptions, the simple average of components commonly reported might not, in general, be such a bad compromise. Sentiment indicators can still be viewed as useful summary statistics complementary to an assessment of current conditions, but the extent to which they augment information already available to us should not be exaggerated. On balance the conclusion is rather disappointing for the supporters of confidence surveys. While it is unclear whether the surveys actually measure that ephemeral concept ‘confidence’, it is clear that, with a couple of exceptions, whatever the surveys do measure does not have much predictive ability. That is, confidence surveys don't appear to tell us much that we didn't already know.


We do not include filtered indices whose underlying regressions are suspected to be structurally unstable. [23]