I looked into this and if you trim only negative outliers the results aren't as good as trimming all outliers. (By this I mean that in-sample returns aren't quite as predictive of out-of-sample returns.) The reason is that all outliers have a truly outsize effect on the slope and intercept of the regression line since least-squares regression is the result of squaring the vertical distance of each point to the line. So points that are particularly far from the line, whether on the negative or positive size, have an outsized effect on the line.