Hieu Tran and Hong Yu train an execution policy on order-by-order replay for 954 Nasdaq stocks from 2019. Their transfer tests leave the learned weights unchanged but allow half an hour of observations from the target to normalize inputs. Giving a raw-feature policy those same observations and a reward measured in spreads recovered 81% of its disadvantage. At the median after fees, the raw features did at least as well. The hard out-of-sample test was not a different ticker: none of the representations handled a previously unseen spread regime.

Sources: arxiv.org