[Journal:] Global Journal of Management and Business Research C [ISSN:] 2249-4588 [Volume:] 17 [Issue:] 4
A well-known issue associated with linear time-series models is the so-called spurious regression problem when the variables are non-stationary. To cure this issue, one usually differences the data first, tests the stationarity of the first differences, and then runs regressions on the revised data. Alternatively, error correction models (ECMs) can be used if the dependent and independent variables are co-integrated. In this paper, we investigate forecasting performance between first-difference models and ECMs through four different simulation designs: (i) non-correlated I(1) processes; (ii) correlated near-stationary I(0) processes; (iii) correlated but un-cointegrated I(1) processes; and (iv) cointegrated I(1) processes. Our results show that ECMs have more robust performance than first-difference models in terms of coefficients estimation and out-of-sample forecasting under the CCAR framework. A simple application for models constructed for banks’ Comprehensive Capital Analysis and Review (CCAR) exercises is exhibited.