A criterion for the number of factors in a data-rich environment

This paper derives a new criterion for the determination of the number of factors in static approximate factor models, that is strongly associated with the scree test. Our criterion looks for the number of eigenvalues for which the difference between adjacent eigenvalue-component number blocks is maximized. Monte Carlo experiments compare the properties of our criterion to the Edge Distribution (ED) estimator of Onatski (2010) and the two eigenvalue ratio estimators of Ahn and Horenstein (2013). Our criterion outperforms the latter two for all sample sizes and the ED estimator of Onatski (2010) for samples up to 300 variables/observation