A take-home asks you to regress 500 daily returns on the Selvara signal. You get a slope of 1.8, but three days with returns above 20% drive the fit. After winsorising at the 1st and 99th percentiles the slope is 0.4. Which do you report and why?

Report the winsorised slope of about 0.4 as the estimate, and show the naive 1.8 and the outlier check beside it. Three days out of 500 set the naive fit, so 1.8 describes those days, not the signal. Before deciding, check whether the three returns are data errors, such as an unadjusted corporate action, and how the slope moves when each is removed. The honest result is a weak relationship, not a strong one.

Ask ten people their monthly income, include one billionaire, and the average describes nobody in the room. Least squares minimises squared errors, so a point far from the rest pulls the line towards itself with a force that grows with the square of its distance; three returns above 20% among days that otherwise stay within 5% dominate the fit. They also sit at high signal values, where a point has the most leverageIn a regression, how far a point sits from the average of the explanatory variable. High-leverage points can swing the slope on their own., so they tilt the whole line. The naive slope of 1.80 is the slope of three days, not of the signal.

First, whether the three days are real. A 22% one-day move in a stock is rare; a stock split or bonus issue that was not adjusted in the price series produces exactly this kind of jump. Look up the three dates: if they are data errors, fix the data and the question disappears; if they are real events, such as a takeover bid, ask whether the signal could plausibly have predicted them. Second, how fragile the estimate is. Rerun the regression without each of the three days and without all three, and try a robust method such as winsorising or a regression that down-weights large errors.

Lead with the robust number and put the evidence for the choice next to it. A research write-up that reports 0.4 and shows that 1.8 came from three days reads as careful; one that reports 1.8 reads as a result that will not survive a live test. Say what the robust estimate means in plain terms: a one-unit move in the signal predicts roughly 0.4 percentage points of next-day return on an ordinary day, with a lot of noise around it: its t-statistic of about 1.7 is weak evidence on its own. Then say what you would do next: test the signal on a later period it has not seen, because a relationship estimated on one sample, however carefully, is still one sample.

Say the limitation of winsorising too. It is a choice, and the 1st and 99th percentiles are a convention, not a law. If the three days were real and the signal genuinely warns of jumps, capping them throws away the most valuable information in the data set, so the right conclusion is not that the jumps do not matter but that three examples cannot prove the signal predicts them. The right amount of care is modest: check the data, choose the estimate, show the evidence, and move on to testing the signal.