Did opinion polls display herd behaviour?

Academic research on analyst forecasts of corporate earnings provides interesting insights, which can explain the failure of almost all polls (except Today’s Chanakya) to come out with accurate seat forecasts for the 16th Lok Sabha.  Brett Trueman, Professor of accounting at the University of California, Berkely shows that financial analysts display a tendency to release forecasts which are closer to general expectations.

They also tend to herd, whereby the analysts make sure that their forecasts are not drastically different from the forecasts released by others. The above mentioned phenomena play out even when the superior private information that the analysts has warrants a different prediction. It is important to note that the above behaviour only reflects a tendency to play safe and confirm to storytelling rather any other ulterior motive such as corruption.

 

 

A closer look at the recently announced opinion polls and exit polls reveals the following: First, other than Today's Chanakya, almost all polls clustered within a range of 260-280 seats for the NDA and 90-110 seats for the UPA. Second, remarkably, vote share predictions turned out to be far more accurate than seat predictions.  Third, state level projections clearly showed a tendency to confirm to the average. Finally, markets were not surprised drastically even though elections threw up a single party majority government after 30 years, a best case scenario for markets. 

 

These findings are perfectly in line with Trueman’s postulates. One of the pollsters confessed that they tend to be conservative in projecting seats for the winners. State level analysis makes this point clearer. Those polls that gave very high number of seats to BJP in Karnataka, AIADMK in Tamilnadu and  BJD in Orrissa in the initial rounds of opinion polls, started bringing their projections down at later stages. It is possible that the pollsters who made correct predictions at the beginning chickened out and showed a tendency to confirm to the average. Trueman shows that markets are likely to factor this aspect into account while reacting to earnings forecasts and announcements. Hence the reaction on May 16th was rather muted.

 

Daniel Kanheman, the nobel winning economist and psychologist, writes in his famous book “Thinking fast and Slow” that human mind is an associative machine, which constantly looks for stories to explain any phenomenon. System 1, the part of our brain which does  fast intuitive thinking for us, is lazy and bases its judgment mostly on story telling rather than rational analysis.

With this background let’s take the case of Karnataka.  Pollsters had a readymade story that the voters are likely to dislike the entry of Yediyurappa and Sriramulu,  given the anti-corruption sentiment  prevailing in Karnataka (Unfortunately no one even bothered to check whether Sriramulu has any criminal case pending against him).This justified a low seat prediction for BJP. These kind of stories might have raised question marks in the minds of those pollsters who started out with 18-20 seats for BJP. Many of them started revising their forecasts downwards. 

 

The case of Orrissa is even more interesting.After Narendra Modi started attacking the performance of the incumbent government, everyone got a perfect story which said that such an attack is likely to position BJP as a challenger to Naveen Patnaik and lead a lot of voters to vote for the BJP. This lead to a series of re-adjustments with respect to BJD’s seat share. A closer look at predictions made for Tamilnadu and West Bengal reveal a similar pattern.

 

It is possible that a pollster is torn between statistical analysis and conventional knowledge of "ground realities". Unfortunately many of them seem to have gone with the latter. In cases where there was no conflict between the two, their predictions turned out to be accurate. Gujarat is a case in point. Everyone had a story that a state whose leader is a Prime Ministerial nominee, is likely to get support. Similar is the case with Rajasthan, Madhya Pradesh and Chattisgarh. There were equally compelling stories.

More importantly, there were no opposing stories on the ground. However, stories emerging from UP and Bihar were not unidirectional. While there was an agreement regarding existence of Modi wave in these states, there were also counter stories such as polarization of minority votes and resurgence of RJD chief Lalu Yadav.These counter stories would have forced pollsters to temper at least seat projections. Everyone agreed that nearly 40% vote share in a multipolar election points towards a sweep but not one other than Chanakya had the conviction to project more than 56 seats for the victors in Uttar Pradesh.

 

The above analysis has highlighted both positive and negative aspects of opinion and exit polls in India. They deserve a huge compliment for getting the vote share projections directionally right. This is not a minor achievement given India’s diversity. This also points out that the ground level personnel of all opinion polls have done a commendable job. It is also heartening to note that sampling seems to be robust. However while converting the vote share projections to seats,  it is likely that the pollsters have fallen prey to herding and storytelling. These results clearly show that psephologists will do well by sticking to statistical modeling rather than falling prey to stories told by ground level “experts".

 

(The author works for Center For Analytical Finance, Indian School of Business)