Hi, I got this problem that I'm trying to see whether there is literature is out there to attempt to solve the problem. So like most of quant strategies, you will have multiple alphas.. like mean reversion, momentum, fundamental, etc.. How do you guys combine them? (Obviously it has to do with quality of the alpha, etc.. and I read somewhere about doing optimization on the alphas to take into the account of the sd of the alphas). But what about rule based stuff? Like today is day after earnings.. so you want to kick in your earnings signal but ignore all the other signals.. and I don't want to put a bunch of the if else statement and blindly run sims... Is there a better way to solve the problem without doing an optimization on the alpha itself? There's regression tree, bayesian, non linear etc.. but can't seem to find that target combining alpha.. How do you guys generally solve this issue?

This seems really a good question. I would like to hear from any advice too

surely you cant have too much alpha ?

You could do a multiple regression but of course there may be problems with multicollinearity. One could perform a factor rotation before performing the regression so you end up using orthogonal factors. Regression weights could be applied across time or some other dimension (e.g. inverse of stock vol, market cap). I'd say the lower the Sharpe ratio of the underlying factors the longer the regression window should be. There's another post in the technical forum that suggests Bayesian regressions.Curious to hear what others have to add to this.

I used to work at one of the big places where they do this. All alphas were normalised to (0,1) and then combined by a simple weighted sum. Weights were subjective (and tended towards being equalish across the 'uncorrelated' themes)They defended the perceived over-simplicity of this through statements such as:1 - RELATIVE historical alpha performance is a poor indicator of future relative performance2 - Factor timing is difficult (so much so that we shouldn't use it)3 - We mostly sold a constant risk productI could buy 1 and 2, but ignoring information in the distribution of alphas really sucked. Imagine a situation where stocks tend to be valued about right and another situation where valuations are widely dispersed... I would rather take a bigger bet on valuations when there is some dispersion (which we expect to close).

if this is genuine alpha then you can't have too much of it...