Methodology
Every BUY or SELL is compared with the same instrument's price five
trading days later; BUY is correct if the price rose, SELL if it fell. HOLD and WATCH carry no
directional claim and are not scored. Edge is accuracy minus the
baseline — the share of all instrument-days in the same
opportunity set that rose, so the comparison is against picking at random from the same
universe rather than against our own picks. Intervals are 95% and are computed on the naive
binomial assumption; because agents evaluate overlapping windows on the same instruments, the
true intervals are somewhat wider than shown.