How to distribute weights in the rank.

I would like to know your opinions about the weight distribution in ranks.

For instance, the screenshot below is the weight distribution of my last ranking. The ranking shows better performance with this weight distribution, three nodes with their subnodes weighted more than the rest.

But then I replicated the same ranking subtracting all the nodes with less weight.

And the performance of that last ranking, was very similar to the previous one.

Now I have few doubts:

Giving a lot of weight to a certain node/s invalidates the rest of the nodes?

If I do this (Give a lot of weight to a certain node in order to achieve more performance in the sim.), am I inside the overfitting zone?

It’s better to have more balanced or similar weights in the ranks?

Thanks.


1.png


2.png

I believe that this is ultimately up to each user. I spend a lot of time optimizing rank weights by trying various combinations on different universes and backtesting the hell out of them. But in the end I believe it makes little difference in the long run. And yes, this does fall under the category of “curve-fitting,” though there are robustness tests you can do to mitigate that. It all depends on what you feel comfortable with and how much time you want to spend on it.

There’s no good argument for equal weights except that it’s easier than assigning unequal weights. The reason I say that is that factors are often correlated. If I have two value factors and a growth factor and assign them equal weights, the result will be very different than if I have two growth factors and a value factor.

I always prefer more factors to fewer factors. I believe that type II errors (NOT using a factor that’s important) are more dangerous than type I errors (using a factor that’s useless or detrimental). When I use ranking systems, I’m trying to evaluate every stock I buy from as many different angles as I can. I’m trying to be thorough. So the more factors, the better.

I hope this helps, as vague as it is.

So I agree with every bit of this except that I may not use as many factors. But I do not know the correct number of factors.

What Yuval says about different weights for different factors can be shown formally with Factor Analysis which uses what Yuval points to: correlations among the factors. So I agree completely on this and for the reason that Yuval gives here. And it can probably be proven for stationary factors (and linear methods).

But for sure correlation among factors can give problems. There is a large body of work in this: usually called multicollinearity in the literature. Marc Gerstein has discussed this too. So no one would argue that Yuval is not discussing an important point here.

Of course, my analysis (using factor analysis) kind of assumes that the market is stationary (the effect of factors not changing much over time). But even if this is a bad assumption that would not go far toward proving that equal weights are best.

Jim

Sincerity always helps. Thanks Yuval. :slight_smile: