What would be the best solution for converting an AI factor's feature importance using the coefficients into a ranking system?
I gave Gemini my coefficient overview, the ranking system (or factor list which is taken from a ranking system) and an explanation of how the ranking system is structured here: https://www.portfolio123.com/doc/RankingDoc.pdf
The best suggestion for how to convert an AI factor into a ranking system was: (se below)
But is it? Are there better ways to do this?
Step 1: Normalize Coefficients into Percentage Weights
Since Portfolio123 combines factors using a weighted average that must total 100%, you need to convert your raw AI coefficients into percentages.
-
Add up all the coefficients in your list to get a "Total Sum".
-
For each factor, calculate its percentage weight: (Coefficient / Total Sum) * 100.
-
Enter this resulting percentage as the "Weight" for that specific node in Portfolio123.
Step 2: Prune the "Noise" (Cut down the number of factors)
A ranking system with 100+ factors will be slow, hard to maintain, and highly prone to overfitting.
-
Keep only the top factors that make up about 75–80% of the total cumulative weight.
-
Discard the factors at the bottom of your list, as their weights are so mathematically insignificant that they won't noticeably impact the final ranking anyway.
Step 3: Define the Direction (Higher vs. Lower is Better) – Critical!
The AI feature importances only tell you how important a factor is, not which direction it points.
-
You must manually evaluate each factor you kept and assign the correct logical direction in Portfolio123.
-
For example: CF/A (Cash flow) should be set to Higher is better, while Volatility should likely be set to Lower is better.
Step 4: Set "NA Handling" to Neutral
By default, Portfolio123 severely penalizes companies with missing data (NA), pushing them to the bottom of the rank. AI models, on the other hand, usually treat missing data neutrally or assign it to an average.