Every result sorted into five equal groups by its score, lowest on the left. Each bar is that group's average return against the Nifty over the next three months. A score that works climbs from left to right.
Against the Nifty 50, and against the average company that reported in the same season. The second is the fairer test: when small companies rally, almost everything beats the Nifty, and comparing with other results takes that out.
Rank correlation runs from −1 to +1: above zero means results that scored higher on this signal tended to do better over the next three months. In stock returns, anything above 0.05 is a useful signal. The weight is how much the tuned score leans on it.
Each season is scored by a model that only learned from results whose three-month outcome was already known before that season began, the way it would have worked live.