Journals are, after all, a filter for high-correlation results using small data sets. By definition that includes every coincidental result out there. You can probably create a distribution of accuracy with enough historical data.