The paper, posted on arXiv, challenges the standard way of comparing time series forecasting models. It argues that pointwise error, which scores a prediction in isolation from the decision it is meant to support, is not a reliable proxy for downstream performance.
In the context of stock return prediction, the authors contend that a lower forecast error does not imply a better decision. Trading profit, not forecast accuracy, is the metric that matters for the decision.
The paper's approach is to evolve small recurrent networks for stock return prediction. The source provides only the opening of the abstract, so no experimental results or conclusions are available here.