A very simple method, but could this work:

[Whycliffes]

I have downloaded my factors and have the following overview with normalized numbers for each factor (feature or nodes):

In column D, I have three months of future relative returns compared to SPY( FutureRel%Chg_D(65, GetSeries("SPY"))).

What if I

1. Filtered out the 5,000 best performers from a total of over 100,000 lines in column D.
2. Then, I add a new row nr. 1, that sums the number of normalized values that are 1 for each factor (feature). (I may have misunderstood this, but 1 is the highest score each factor can achieve.)
3. Then, I create a ranking system based on, for example, the 40 nodes that have the highest number of 1s among the 5,000 best performers.

Wouldn't that provide a very good indication of which of my 300 features (nodes) perform best in relation to future three-month returns?

The idea here is that the features with the largest amount of 1 in the 5000 best performers give a indication of what characteristics that would be a part of the winning stocks.

[ZGWZ]

It is better to perform this operation in reverse for 5,000 worst performers

[Whycliffes]

Why? And how would you do it?

[ZGWZ]

AI-linked companies are now >10% of global market cap

Summarising from positive examples is less effective for the same reason as stated earlier: stocks that go up big can be winning lottery tickets (like GME and TSLA) or they can be really good stocks. However, stocks that go down big generally do have serious problems and are therefore better suited to get useful results. For example, it is more useful to draw lessons from the failures of SGMA than from the successes of NVDA. Of course, I'm actually not prepared to analyse any individual stocks…

[euroventure]

@Whycliffes

FutureRel%Chg_D(65, GetSeries("SPY"))).

fantastic insights, I shall be bashing the bishop all day over those insights

[Whycliffes]

Thank you, but the idea came in a conversation with Test_user. Keep us updated on how your testing goes..

[Whycliffes]

But would a large enough sample base—even more than 5000 help to model away from fitting the strategy to the "winning lottery tickets"?

One question, when you normalize numbers, am I correct in understanding that 1 is the highest. So if the stock factor (node) is a measure of high insider buying, the largest amount insider buying in the sample set will get a normalized value of 1?

[ZGWZ]

As I recall, the machine learning model trained on the 1/5th of the stocks with the highest future returns in the paper I'm talking about got OK returns, but they were extremely volatile, and the return-to-risk ratio was not even as good as the model trained on the 1/5th of the stocks with the second highest future returns.

[Whycliffes]

"This is the 50 features with the highest number of '1's. Equal balanced.

Not the best result:"""