For future elastic optical networks, the narrow filtering effect induced by cascaded reconfigurable optical add–drop multiplexers (ROADMs) is one of the major impairments. It is essential to accurately estimate the filtering penalty to minimize network margins and optimize resource utilization. We present a method for estimating filtering penalty using machine learning (ML). First, we investigate the impact of ROADM location distribution and bandwidth allocation on the narrow filtering effect. Afterward, an ML-aided approach is proposed to estimate the filtering penalty under various link conditions. Extensive simulations with 9600 links are implemented to demonstrate the superior performance of the proposed scheme.