The Atlanta Fed's macroblog provides commentary and analysis on economic topics including monetary policy, macroeconomic developments, inflation, labor economics, and financial issues.

Authors for macroblog are Dave Altig, John Robertson, and other Atlanta Fed economists and researchers.

« Another view of the structural versus cyclical unemployment question | Main | Increasing hours worked versus increasing hiring »

June 18, 2010

Another look at consumer sentiment and consumer spending

In the most recent economic forecasting survey by the Wall Street Journal, 23 percent of the surveyed economists said consumers spending more readily than anticipated is the biggest upside risk of their growth forecast for the second half of the year. So anything that can shed light on future spending habits is of particular interest. Two of the most commonly cited measures of consumer attitudes are the Conference Board's Consumer Confidence Index and the Thomson Reuters/University of Michigan's Index of Consumer Sentiment. A key question is, do these indicators improve consumption forecasts?

Previously, economic researchers have looked at the predictive power of these indexes for consumer spending, and they generally found that the ability of consumer confidence measures to predict consumer spending largely disappeared once some other measures of economic conditions were taken into account. One such example is a study by Sydney Ludvigson, which examined the forecasting record of these confidence measures through 2002 (for other examples, see here and here). Much has happened since then, of course, and a simple inspection of the two series reveals that both confidence measures fell fairly steadily starting in August 2007 until reaching near-record lows by June 2008. Therefore, a look at the more recent predictive track record of these indicators seems warranted.


For this examination, we conducted an out-of-sample forecasting experiment using a pair of statistical models (technically, Bayesian vector autoregression models). The first model predicts real personal consumption expenditures as a function of its own past values and past values of other variables such as real measures of stock market prices and disposable personal income. The second model includes all of these variables augmented by the two measures of consumer attitudes. At each point in time we use only the data that would have been available to forecast real consumption data anywhere from one to 12 months out. (For example, in the middle of February 2009, consumption data would have been available through December 2008 while some of the other variables would have been available through January or February 2009. The experiment is not "real time" in the sense that we use the latest vintage of data, which include revisions to the historical data that would not have been available to forecasters at the time.) Forecasts of consumption are made for the 1990–2003 period and then again for the period from 2004 to the present. The root mean squared forecast error is used to gauge the accuracy of the forecasts, with smaller numbers corresponding to smaller misses on average. As the accompanying chart shows, adding the two measures of consumer attitudes improves the forecast much more in the post-2003 sample than in the earlier period.


We experimented with some variations in specifications of the model, and we were unable to overturn the general finding that adding attitude measures to the model resulted in an improvement in forecasts in recent years. We found this fact intriguing and somewhat surprising.

A recent paper by Barsky and Solon argues that the Index of Consumer Sentiment reflects the public's awareness of economic conditions. In fact, the survey used to construct this index asks respondents about recent news they have heard related to changes in economic conditions. From August 2007 to June 2008, news of "unfavorable higher prices" was frequently mentioned in the survey. A study by James Hamilton showed that part of the deterioration in the Index of Consumer Sentiment during this period could be explained by rising energy prices. However, adding a measure of oil prices to our model did not overturn the basic finding of improved consumption spending forecasts in models that included measures of consumer attitudes.

It remains an open question why these measures of consumer attitudes have become more useful in recent years. A statistical anomaly, greater or more accessible news coverage of the economy, and a generally more aware public are all possibilities. If it is just luck, then time will eventually overturn the result. But if these consumer attitude indicators have become a more useful summary of a wide variety of developments in the economy, then their forecasting power will persist. Time and further research will help sort this out.

By Patrick Higgins, an economist at the Atlanta Fed

June 18, 2010 in Data Releases , Forecasts | Permalink


TrackBack URL for this entry:

Listed below are links to blogs that reference Another look at consumer sentiment and consumer spending :


Very interesting quantitative analysis here. It would be interesting to see an attempt to quantify media news and the impact it has on consumer confidence. Probably wouldn't be much help in forecasting, though.

Posted by: Robert F. | June 23, 2010 at 10:50 AM

I don't understand this graph. Why does forecast error decrease with horizon? Shouldn't it be easier to predict the immediate future?

Also, can you give an impression about number of data points v. number of parameters?

Posted by: Doug | June 28, 2010 at 05:47 PM

consumer sentiment will always drop out once employment and income enter the equation -- this was true 30 years ago and it still is today. where current econ environment can lead a spending equation astray is if the household balance sheet, net worth / dpi, is left out. reason why income growth in current cycle will bring about more saving than strict income/employment equation would suggest. perhaps here is where some part of the sentiment survey can work as a proxy for wealth.

Posted by: steve | June 30, 2010 at 09:32 AM

Patrick Higgins responds: The forecasting experiment for this blog post uses monthly data starting in 1968 and ending anywhere from 1989 to 2009, depending on when the forecast is made. A particular forecast would include anywhere from about 250 observations to about 500 observations. The basic model includes 13 lags of eight variables plus a constant term for each variable, implying it has a total of 112 parameters. The model that includes the two measures of consumer confidence has 140 parameters. The forecasts are out-of-sample (i.e., they don't use future data the forecaster wouldn't have at the time), so the larger model does not necessarily have an a priori advantage over the basic model.

The accuracy of the forecasts are gauged using errors in predicting annualized growth rates--measured by logarithmic changes--in consumption anywhere from one month to 12 months in the future. As can be seen in the graph below when growth rates are annualized, the one-month growth rate in consumption is quite noisy, while the one-year growth rate is fairly smooth, and the 10-year growth rate is smoother still. This activity is essentially a particular case of the law of averages.

Chart 1

Therefore, one might expect that annualized growth rates become easier to forecast the further out in the future one goes. This case turns out to be one of using the simplest forecasting models of consumption growth--the so-called "naive model," which projects that annualized consumption growth in the future will be equal to its historical average (since 1968, in this case)--and is pictured in the graph below.

Chart 2

Posted by: Patrick Higgins | July 02, 2010 at 12:53 PM

Post a comment

Comments are moderated and will not appear until the moderator has approved them.

If you have a TypeKey or TypePad account, please Sign in

Google Search

Recent Posts



Powered by TypePad