An artificial correlation that can arise when too many properties are screened relative to the number of available observations.
Example: If one tests 20 possible random descriptors for statistical significance in a multiple regression equation of the properties of 15 compounds, the average fitted \(R^{2}\) is \(\pu{0.81}\) even though an average of only four descriptors were included in the equation.
Source:
PAC, 2016, 88, 239. 'Glossary of terms used in computational drug design, part II (IUPAC Recommendations 2015)' on page 243 (https://doi.org/10.1515/pac-2012-1204)