LIBOR meets machine learning: A Lasso regression approach to detecting data irregularities

Abstract

We use the Lasso linear regression technique to detect level shifts and additive outliers in both the daily 3 and 6 month US dollar LIBOR rates, and compare the results to an identical application of this technique to non-LIBOR, US short-term funding benchmarks. We find that the two LIBORs have the largest incidence of outliers, especially, additive outliers. For two non-LIBOR benchmarks, the 6 month Treasury bill and the Federal funds effective rate, no outliers were detected, which reinforces our results. Furthermore, our identified outlier episodes for both LIBOR rates fall inside the period that the manipulation of LIBOR occurred.