Today we are releasing Dataset v3.0 “Supernova” dataset, which is ready immediately. Just over 1 year ago, we released the [risk-aware scoring system for Signals]( Introducing Alpha: A Risk-Aware Scoring System for Signals ) to mitigate one of the most difficult parts of Signals: drawdowns in scoring. This release further improves Signals by adding newly-identified risks to the neutralization matrix and the new target is better than ever, making this release the most important upgrade to Signals yet.
Features and Targets
The feature set is the same as previous versions. However, the target is different and the 2 major supporting files have drastically changed: new neutralizers and no sample weights.
In the v2.1 “Alpha” release, we created a neutralizer matrix and a related sample weights vector. These were used to make Signals scoring risk-aware, but we’ve since integrated the sample-weights directly into the new “jupiter” target.
Without the sample weight vector, a few other major changes follow:
-
The neutralizers are simple enough to use our existing “neutralize” function.
-
The “alpha” score is replaced with a “neutral correlation” score against the new target.
-
Instead of “meta-portfolio contribution”, we will start using a simplified “neutral contribution” function against a new v3 Neutralized-User Stake-Weighted Meta Model and the new target.
-
The “turnover threshold” is being replaced by a “neutral churn penalty” described below.
Here is a breakdown of the release timeline:
-
Signals Data v3.0 is ready today (August 28, 2026).
-
The turnover threshold is being deactivated today (August 28, 2026).
-
The new scores and neutral churn penalty will be displayed on the website by September 4, 2026.
-
Payouts for rounds starting on or after September 25, 2026 will switch to using “neutral correlation”, “neutral contribution”, and the “neutral churn penalty”.
Neutral Churn Penalty
The turnover threshold is being replaced with a generalized payout penalty, limiting upside for untradeable signals. Instead of blocking your stake, we will allow you to continue staking, but payouts will diminish proportional to the churn of your neutralized predictions.
For each submission, we will neutralize it and calculate the churn with respect to previous neutral predictions. The mean neutral churn across the last week of submissions will be used in the penalty calculation below:
Here, x is the mean weekly neutral churn we just calculated. The “10” is a constant scaling factor and the “0.1” is the churn threshold. Going over the churn threshold will diminish payouts proportionately:
At a neutral churn of 0.2, payouts are diminished by 50%, and at a neutral churn of 0.5, the predictions are effectively un-tradeable, so this kind of signal will not be paid very much. This churn penalty goes into effect for rounds starting on or after September 25th.
Benchmark Models
The existing Signals benchmark model was retrained on Jupiter and shows significantly improved neutral correlation against the jupiter_60 target and reducing max drawdown in neutral correlation:
Start neutralizing your signals today.
Happy Modeling