Many AI factor upgrades, plus 50% discount on training

Dear All,

Many new upgrades were released for AI Factors. Each deserves it's own post and knowledge base article, but I'll start by listing them first. Please ask us for clarifications.

50% discount on training

First of all, we're discounting training costs by 50% for the next month to entice you to try all the new features. We're also planning a fixed monthly cost tier with unlimited training & prediction. Let us know what you think. The training costs you see now already reflect the discount, for example Extra30 used to cost $1.50/h, it is now $0.75/h.

New normalizations options

You can now normalize each feature independently, as well as easily add different flavors of the same metric. For example, for something like Earnings Yield, you can now normalize it by:

  • Comparing to other stocks
  • Relative to past values
  • Relative to Sector or Sub-Sector

It's all too much to list here. We tried to make it all intuitive, but the complexity has definitely gone up. We'll be doing write ups as well as testing everything ourselves to see what works.

Another thing, we are starting to believe that "feature engineering" is much more important than, for example, hyper parameter tuning. Therefore we are working on tools right now to help you test features and normalizations quickly, and select the best set, with the lowest correlations between them. Stay tuned.

New predefined features

We have revamped the predefined features in light of the new normalizations. The organization of the pre-defined features should also make more sense and be easier to navigate.

New predefined models

We added several different predefined model variations. New models adopt a new naming scheme that it's still undergoing last minute changes. Sorry about this, please bear with us. This turned out to be much harder than we thought.

Grid search

You can now use the grid search functionality to quickly add up to 200 different hyper parameter permutations

That's all the upgrades in a few short sentences. More help/docs coming soon.

Thanks

4 Likes

Agree!

3 Likes

Great - The lack of proper feature engineering (and related analytics) was the reason I did not sign up for AI factors.

I also suggest to incorporate composite predictors as collinearity is not a problem when "summing up" in a classic ranking system however it is very much an issue if you plug in 10 different FCF factors in any of the algos on offer.