In recent days numerical models like Wavewatch are used for grid based wave forecast all over the world Considering the resolution it is difficult to scale down a location specific based forecast. As a result univariate wave forecasting may be employed where in previous values of waves are used to forecast wave up to few hours to few days in advance. However such models suffer the vital issue of ‘removal of phase lag’ which has been recognized by many researchers and attributed to high autocorrelation between last two observed values in univariate time series modelling. Authors have successfully removed the ‘phase lag’ in wave forecasting by employing Multilevel Neuro-Wavelet Transform. This is an extension of that work which targets towards exploring the behavioural aspects of different decomposition levels of wavelet in wave forecasting at one location along USA coastline in a view to improve the accuracy of wave forecasts at different lead times. The hybrid Multilevel Neuro-Wavelet Transform used in the present work is combination of discrete wavelet transform (DWT) and artificial neural network (ANN). The discrete wavelets analyze frequency of signal with respect to time at different scales and decompose it into low (approximate) and high (detail) frequency components. In the present work the decomposition is done up to seventh level starting from the first (1st, 3rd, 5th and 7th) level. The results were judged by phase angle, phase difference and extreme value predictions along with correlation coefficients rather than with traditional error measures.