Row-position masks are not allowed

Hi fellow AlphaNovans,

We have found a number of Cycle 2 submissions that multiply their predictions by a schedule read from the row index alone: the prediction is set to zero, or scaled down to a tiny opposite-signed residual, on fixed stretches of the history, and left untouched on the rest. The schedule reads no feature, no target and no return. Its only effect is on how the signal is measured. Rows with no prediction contribute nothing to the correlation used for quality-set admission, so several copies of one model can be admitted as if they were different signals, and a signal can sit out stretches of the history, including the live window, while its full-history Sharpe decides its rank.

This is not a modelling choice. It is a way of defeating the admission rule, and it takes quality-set slots and prize money from scientists who submit one signal per idea. We are notifying the accounts involved and disqualifying the affected signals from Cycle 2. The disqualifications will change the Cycle 2 quality set and its provisional winners.

Going forward, any submission whose output is gated or scaled by row position will be disqualified, and the accounts involved will lose prize eligibility for the cycle. We are also tightening the admission check so that a masked copy of an existing signal no longer reads as novel.

A signal that goes flat because the model sees nothing to trade is fine. A signal that goes flat because a hardwired rule says which rows are on and which are off is a mask, and it will be disqualified. If you have submitted something like this, please do not submit it again.

Best,
NONIUS