Alpha

One of the primary scores of Signals using the 60D2L chili target.

CORR, FNC, and MMC are all imperfect metrics for paying Signals users. This is because users can be good at these scores, but backtest poorly. To better align the Numerai Signals tournament with the Numerai hedge fund, the performance of a submission must align with it's backtest performance.

The Numerai research team invented 2 pieces of data that provides a way to do this:

• Neutralizers Matrix: a 200-column condensed matrix used to produce a neutralize signal

• Sample Weights Vector: a weight vector used for sampling a signal to produce a set of stock weights that accounts for real-world trading constraints such as adv

Using these, we can create a set of "neutralized weights" from a signal as follows:

Given the following data:

• signal s (NaNs filled w/ 0.5)

• neutralization matrix N

• weight vector v

We first create normalized signal s’:

s` = normalize(rank(s))^1.5

Then generate neutralize predictions and sample them:

neutral_preds = s` - (N @ (N.T @ (v * s`)))
neutral_weights = neutral_preds * v

Finally, we use the 60D2L target chili to derive the final alpha value:

alpha = neutral_weights @ target_chili_60

This value provides us some insight into how a signal might perform in a backtest and more closely aligns the Signals Tournament with the hedge fund.