Figure 1

Comparison of the cumulative return of the adaptive trading strategy against the baseline buy and hold strategy and three benchmark strategies in a backtesting experiment running from January 6th, 2008 to March 26th, 2017. The experimental period is split into a validation period, used to tune the hyperparameters of our model, and a testing period. The baseline buy and hold strategy simply follows the market by buying the DJIA at the beginning of the experiment and holding the portfolio until the end. Kristoufek’s strategy follows the market with a mechanism allowing it to diversify risk using web search data, leading to a better performance than buy and hold. All the other strategies automatically long or short the whole portfolio or individual stocks on a weekly basis by analyzing web search data. The strategies proposed by Preis et al. and Heiberger are static; they use pre-defined search terms and fixed decision heuristics. Returns for these two strategies saturate after 2012, evidencing loss of predictability. By selecting different search terms and retraining a predictive model for every decision, the adaptive strategy increases the portfolio value by nearly 500% at the end of the experiment, which is 404% more than the most-profitable benchmark strategy (Kristoufek’s).