Figure 2

Effect of changing time window on predictability, measured as the probability of being outperformed by chance (lower probability implies higher predictability). We use the approach described in the caption of Table 2, but compute the probability in six overlapping 4-year windows instead of the whole time span of the experiment (9 years 3 months). The predictability of the benchmark strategies varies as the window moves, whereas the adaptive strategy has consistently high predictability.