Correlation of Democracy Indicators and Markets Returns

Abstract

We perform various experiments correlating past changes of social indicators about a country with future stock market returns for that country. The 169 social indicators we use, which go back as far as the year 1900, are available from the Varieties of Democracy Project. We use two sets of data for country-wide stock market returns: data compiled by Dimson, Marsh, and Staunton covers 17 countries going back to 1900, and data from the MSCI data analytics and index service covering 45 countries going back as far as 1970. We consider five and ten year time windows. This gives us four different “studies”: MSCI 10 year, DMS 10 year, MSCI 5 year, and DMS 5 year. We find the striking result that good changes of the social indicators have a positive mean (averaged over studies) total correlation (correlation of change vectors indexed by country-year pairs) with future stock market returns in 157 out of 158 cases in which the indicator measures something good or bad for society. We obtain a result almost as strong when the correlation is aggregated differently using the separate country and year groupings. We perform statistical hypothesis testing to show that, even though the social indicators are not all independent, these result are exceedingly unlikely to be the result of random (white noise) stock market returns. We also perform “positive linear regression” of stock market return on all 158 indicators, which means that the sign of the regression coefficient for an indicator is constrained to be positive or negative according to whether a positive change of the indicator is good or bad. The fraction of data explained by positive regression is shown to be extremely statistically significant. We calculate a confidence interval for the percentage of data genuinely explained by regression, not just by fitting to noise. The lower end of the confidence window for the four studies is 11%, 14%, 6%, and 9%. We include a long appendix on the statistical theory of correlation and (unconstrained) regression. This provides background to the novel applications of hypothesis testing and confidence interval calculation in the body of the paper.