In a recent Live Stream, Numerai released the historical values of the Signals Naive Meta Model. This is a Meta Model which is built by equal weighting all the signals on Numerai Signals (https://signals.numer.ai) with a stake >0.1. Watch the live stream for more details.
Because Signals is all about submitting predictions on raw stock tickers, you can now see the raw values of the Signals Meta Model with a 3 month lag. You can download these simply by signing into Signals and clicking Data to download all data.
I looked at the correlation to some basic Numerai features myself.
Here are 10 features the Signals Naive Meta Model has a lot of correlation with on a recent date:
Feature, Correlation with Signals MM
Williams %R Indicator, 0.518934
Forecast Earnings Volatility, 0.362848
Volatility of Volume Divided by Price, 0.344192
Residual Volatility, 0.318293
Earnings Dispersion, 0.256236
USD Market Cap, -0.444220
Bollinger Bands, -0.494701
Commodity Channel Index, -0.509648
Stock Price (Ranked By Country), -0.535897
Price to 52 Week High, -0.628304
First of all, you can see there is a lot of correlation with technical indicators (emphasized in italics) on the long and the short side. These features are easy to produce with price data. And since Numerai doesn’t give out data for Signals, many users seem to be using price data in their signals.
The problem with technical indicators such as these is that they are very well known among market participants, and they also tend to be very high churn. A stock could have attractive Williams %R Indicator score in one month and then a very bad score a month later. The problem with this is that trading costs make this signal unmonetizable in a large hedge fund. The trading costs and market impact costs required to trade in an out of the stock quickly would remove the edge.
I would encourage Signals users to try to reduce their reliance on high churn technical features. This will improve your ability to earn TC and decorrelate you from the other signals on Numerai. You can see your churn in Diagnostics.
Another thing you’ll notice is that the Signals Meta Model is heavily short high market cap stocks. It wants to go long small caps and short large caps. Since Numerai neutralizes signals to size before scoring them, taking size exposure like this is unlikely to be rewarded. I don’t think it makes sense for anyone to submit a signal correlated with size or (similar) the stock price.
The Price to 52W High exposure seems especially large (-0.62 correlation). The Signals Naive Meta Model would have a tough time performing well if that feature were to reverse. New models which reduce the exposure to that would probably be very additive here.
I hope this is useful to the Numerai Signals users. And of course, please not there could be many more features which you are correlated with which Numerai can’t detect from our features. This is indeed the point of Signals and since Signals is additive to the Numerai Meta Model it is very likely that many Signals are contributing powerful signals we can’t detect with a simple correlation like this.
Given that the Signals Meta Model produced by hundreds of models is now public, I’m curious to know what other feature correlations can be detected in the signal and what else you can see in it.