First impressions/questions on AI Factors Beta

[mike0001]

Or does Port123 staff envision a ranking system that utilizes several combined
predictors

[feldy]

@mike0001

Has anyone suggested creating an option where factors/feature can only have a maximum weight in the prediction say 5-10% to help reduce curve fitting or protect against changing factor regime? Or better yet where the user can enter a maximum weight for instance in a 80 feature dataset the user could enter a maximum weight value of say 5%. While this may reduce back tested results it may result in better live port returns.

Limiting the feature weight would really be the domain of the underlying machine learning algorithm, unless you were to do some manual postprocessing of the model parameters after it's been fit.

If you're doing a linear regression, p123 exposes both lasso and ridge regression. Lasso uses L1 regularization to penalize the model on the sum of the absolute coefficients, so it can help steer the model towards zero coefficients for sparsity in your model. Ridge uses L2 regularization which penalizes the model on the sum of the square of coefficients, so it can avoid too large of a coefficient on any one feature.

If you're worried about changing factor regime, I think having the model periodically refit would be good strategy, and that's why I'm hoping we can eventually get that support in p123.

@mike0001

Or does Port123 staff envision a ranking system that utilizes several combined
predictors

I'm assuming you're referring to combining several AI predictors, and I absolutely plan on testing this. It's referred to stacking in the machine learning world and is a common technique for improving performance.