The performance of four different cumulus parameterization schemes (CPS) in the Weather Research and Forecasting (WRF) model for simulating three heavy rainfall episodes over the southern peninsular Malaysia during the winter monsoon of 2006/2007 were examined. The modelled rainfall was compared with the 3-hourly satellite observation and objectively scored using a verification technique called the acuity–fidelity. The technique is based on minimization of a cost function that is calculated from four parameters taking into account errors in distance, time, intensity, and missed events. All simulations were made for 72 hours for the three episodes starting at 1200 UTC 17 December 2006, 1200 UTC 24 December 2006 and 1200 UTC 11 January 2007, respectively. The four different CPSs used are the new Kain–Fritsch scheme (KF2), the Betts–Miller–Janjic scheme (BMJ), the Grell–Devenyi ensemble scheme (GD) and the older Kain–Fritsch scheme (KF1). While the BMJ scheme shows some success in the second and third episodes, it shows high location errors in the first episode, leading to high acuity errors. The GD, KF2 and KF1 schemes performed poorly, although both the BMJ and GD schemes simulated the observed drastic increase of rainfall at 2100 UTC 18 December 2006 during the first episode. Overall, the KF1 and KF2 schemes produced positive biases in terms of coverage, while the GD scheme showed persistent location bias, producing a scattered line of precipitation over the eastern coastline of peninsular Malaysia. Although the BMJ scheme has better results, its poor performance for the first episode suggests that suitability of CPS may be case dependent.