In a thread Delaney replied to someone
You're correct that many factors have signal only on the long or short sides, so that's something which is important to investigate.
From the Risk Model example this was my rough start.
The idea at first was to use the top or bottom of the values of one fundamental for long and another for short, then piece them together as input to MaximizeAlpha(). Notice the minus signs, reversals that worked out better. This is just 2016. With the date range of the original, it doesn't do well.
Stocks will have signals for both fundamentals, maybe the strongest could be chosen after ranking or some sort of normalization.
Simply adding them would look at lot like combined_alpha. It's reasonable if equally long and short signals to assign zero weight. Yet in that case long or short could make more sense based on its beta and the direction of the market during the last week or so perhaps.
So how can long and short be isolated?
