A comparative performance of econometric and heuristic forecasts of asset returns

Our project intends to explore the comparative performance of state-of-the-art econometric dynamic methodologies vs. simpler heuristics at forecasting asset returns for important investment categories such as stocks and bonds. Practitioners often report that simpler rules-of-thumb that track current events and rapidly transform news into trading signals often outperform more typical, structured statistical frameworks now popular in empirical finance. However, it remains unclear exactly why, how, and when simple heuristics may actually outperform. In our paper we plan to systematically investigate the sources of differential out-of-sample predictive accuracy of a large set of linear and non-linear forecasting approaches and two broad categories of heuristic frameworks based on the general notions of web-based search frequency and of news coverage.

On the heuristics side, we propose two types of forward-looking indicators: (1) Google search frequencies, and (2) news coverage, proxied by a search for news articles and other broadcasts in news databases. The volume of internet searches and the number of news articles help gauge the degree of investors’ and media’s interest in specific assets, and their time-varying attention to them. Under the assumption that this is caused by increased uncertainty, both searches and media coverage have a potential to forecast subsequent realized asset returns, as recently discussed in the economic psychology literature (e.g., Liemieux and Peterson, 2011).

On the more conventional econometric side, we plan to use a wide range of state-of-the art models, both of linear and non-linear (regime switching predictive regressions, threshold autoregressive) type, admitting in this last case conditional heteroskedasticity effects captured through GARCH models. Predictor variables are those typical of the literature featuring a range of macroeconomic and market leading indicators.

While there is a vast literature that has employed statistical approaches in the attempt to forecast stock and bond 
returns, there is a small but growing literature spanning the empirical finance and business fields that has proposed 
the use of web-based indicators to derive predictive indices. Da et al. (2011), Dzielinski (2012) and Herzog (2012) 
are among the first studies that reported the explanatory power of Google search frequencies over future stock returns.

Finally, our exercises are planned with reference to both US and UK data to avoid the possibility that specific features of the leading role of the U.S. financial markets and media may bias our results in favor of the conjecture that heuristic rules may eventually contain as much predictive power as statistical methods do.