The quality and quantity of stock data is too poor to lead to robust correlations between individual factors.
A systematic search would produce sufficiently (and equally) large numbers of factors in each meaningful domain. Forcing relatively close total weights between long-term and medium-term factors that have a robust uncorrelation with each other also helps (for your mental health, not the Sharpe ratio).
The combination of univariate OLS allows for differences in effect sizes not accounted for by equal weights. However, in practice there is little difference in their predictive effects.
If you are concerned about correlation, do you think PCA or factor analysis could help? As you know, these methods group correlated factors while also making the groups uncorrelated (or orthogonal) to each other through rotation.
I wouldn’t want to recommend how to weight those latent factors for anyone reading this, but they could be equal-weighted, or one could use PCA regression. There are also other weighting methods, some of which I have experimented with.
I’m not necessarily recommending this approach (and I don’t use it now), but I’d be interested in your thoughts.
Certainly correlations between, say, ROI, ROE, ROA, and profit margin are going to be strong because they all have exactly the same denominator. EPS growth and EBITDA growth are also going to be highly correlated. Unlevered free cash flow to EV and free cash flow yield will be highly correlated, as will the ratios of sales to market cap and to EV. All of those are very commonly used factors. So I don't understand why you would say that their correlations aren't going to be robust. I must be missing something here . . .
The kind of correlation you're talking about seems to be able to be utilized by PCA-type methods, but the marignal effect of the best PCA-type method (as IPCA) is even actually basically nonexistent after transaction costs are taken into account. That's why I said it's not useful because of the poor robustness of correlations. Also, they belong to a very small part of the correlation matrix.
Because quite simply, the parts of correlations that are robust are basically the unexploitable parts, and the parts that have a valid amount of information are extremely unrobust.
In my opinion "Factor Momentum" is a valid strategy. Its not easy to develop on here but I have done it. Took me many tears to fully automate it (mostly waiting for P123 functionality to catch up) and uses hundreds of custom formula and ranking systems.
I've been away for a year so haven't been active, all I can says is factor momentum is possible to achieve on P123. There is a thread from a few years ago in which I explain how it works. Can't find it at the moment...
Factor Momentum is different from Factor Weighting, but it does discover Factor Rotation. Factor Weighting comes into play once Factor Momentum is discovered, then the resulting factors are weighted accordingly.