Since initial attempts by Cane et al. around three decades ago, the ability of dynamical models to predict the El Ni？o and the Southern Oscillation (ENSO) has improved significantly. Operational seasonal ENSO predictions now become routine in many major climate centers worldwide by using coupled general circulation models (CGCMs)with sophisticated initialization schemes. On occasions, however, our ENSO predictions can still have notable failures and the forecasts are not yet reliable, two notable examples being the false alarms in 2012 and 2014. In this talk, I will discuss the challenges in ENSO prediction by focusing on initialization uncertainties and model inadequacies. Particularly, I will present results about the role of salinity initialization in ENSO prediction, the contribution of off-equatorial SST anomalies (some signals that could be realistically represented by present CGCMs because of biases in representing thermodynamical process) in 2014 El Ni？o false alarm, and the importance of atmospheric convection parameterization in ENSO prediction.