Flood is a major disaster that happens around the world. It has caused many casualties and massive destruction of property. Estimating the chance of a flood occurring depends on several factors, such as rainfall, the structure and the flow rate of the river. This research used the neural network autoregressive exogenous input (NNARX) model to predict floods. One of the research challenges was to develop accurate models and improve the forecasting model. This research aimed to improve the performance of the neural network model for flood prediction. A new technique was proposed for modelling nonlinear data of flood forecasting using the wavelet decomposition-NNARX approach. This paper discusses the process of identifying the parameters involved to make a forecast as the rainfall value requires the flow rate of the river and its water level. The original data were processed by wavelet decomposition and filtered to generate a new set of data for the NNARX prediction model where the process can be compared. This research compared the performance of the wavelet and the non-wavelet NNARX model. Experimental results showed that the proposed approach had better performance testing results in relation to its counterpart in terms of hourly forecast, with the mean square error (MSE) of 2.0491e-4 m2 compared to 6.1642e-4 m2, respectively. The proposed approach was also studied for long-term forecast up to 5 years, where the obtained MSE was higher, i.e., 0.0016 m2.