Long memory factor model: On estimation of factor memories

This paper considers the estimation of the integration orders of the latent factors in an approximate factor model. Both the common factors and idiosyncratic error terms are potentially nonstationary fractionally integrated processes. We propose a two-stage approach to estimate the factor memories. We show the consistency and asymptotic normality of the proposed estimator. Applying the estimator to the log-squared returns of the US financial institutions, we find evidence of long memory in the estimated factor. We also find that the factor becomes more persistent after 2007.