The Reduced Wisdom of Herded Crowds
Updated: Jan 3, 2020
Often the prediction of a crowd is more accurate than the prediction of any individual in that crowd. Yet, for crowds to be wise, opinions of the individuals making up the crowd must be diverse and independent.
Take the classic jelly beans in the jar guessing game, where contestants have to determine the number of red jelly beans in a jar filled with red and blue beans. For any single participant, it makes sense to extrapolate from the beans they see to estimate the total in the jar. However, the jar may be filled unevenly - such that the front of the jar just happens to have more red jelly beans than the back of the jar. A large crowd that views the jar only from the front will likely overestimate the number of red jelly beans since they share a common biased viewpoint.
We examine a related phenomenon in the equity research industry in our paper, “Are Crowded Crowds Still Wise? Evidence from Financial Analysts' Geographic Diversity.” The equity research industry is geographically concentrated, with over half of analysts working on Wall Street. For retail analysts, store visits are often an important part of their primary research. This practice is widely accepted, with fund manager Peter Lynch once writing “visiting stores and testing products is one of the critical elements of the analyst's job.” Using satellite imagery of retailers’ parking lots we find evidence that the fuller a retailer’s lots were, the better its earnings tended to be.
However, analysts may overweight how representative their local store visits may be. In this sense, the analyst setting is a lot like the jelly beans example, as the consensus of individual analyst forecasts may not eliminate individuals’ errors if the majority of analysts generalize based on the same local information.
A major challenge to testing whether analysts overweight such local information is that such firm-specific local information is typically hard to observe. We overcome this challenge by taking advantage of satellite image data that provides us with parking lot car counts of retail firms across different metropolitan areas at the same time. For example, consider financial analysts working in Minneapolis and Cleveland who each make an earnings forecast for Home Depot. While all analysts could access the same SEC filings and listen to the same conference calls, an analyst in Minneapolis may observe full parking lots at local Home Depot locations and infer that the firm is doing well overall, while the analyst in Cleveland may observe less full parking lots and have a more tempered impression of the firm. Consistent with this intuition, we find evidence that analysts shade their quarterly earnings forecast toward their locally observed parking lot car counts.
While car counts contain valuable information in aggregate, analysts tend to overweight their own forecast in the direction of local car counts - which have no explanatory information beyond the aggregate car counts - relative to other analysts covering the same firm at the same time but in different locations. Our research approach allows us to rule out many competing stories due to common information such as company filings, matching between analysts’ characteristics and the firms that they cover, differences in location characteristics, and time-varying analyst specific characteristics like mood or their general predictions about the macroeconomy.
What implications does this have for the crowd forecast? If analysts make systematic errors that overgeneralize their local information, then we would expect to see the crowd forecast to average away these errors when the consensus is made up of a geographically diverse population - and correlated errors to remain when they are not. Consistent with this prediction, we find evidence that consensus forecast errors are larger when analysts are more concentrated in fewer metropolitan areas (i.e., less geographically diverse).
Our findings provide a more nuanced take on the wisdom of the crowds and point out a potential issue in the market for equity research. With recent consolidation and downsizing in the brokerage industry, individual analysts’ forecasts are becoming less independent and their consensus guidance can suffer as a result. Our results point out one mechanism for why reduced coverage can lead to more uncertainty.
William Gerken - University of Kentucky
Marcus Painter - Saint Louis University