Lorenz time-series is characterized by non-linearity, noise, volatility and is chaotic in nature thus making the process of forecasting cumbersome. The main aim of forecasters is to apply an approach that focuses on improving accuracy in both one-step and multi-step-ahead forecasts. This paper presents an empirical analysis of Lorenz time-series using Scaled UKF-NARX hybrid model to perform one-step and multi-step-ahead forecasts. The proposed hybrid model is trained using Bayesian regulation algorithm. The experimental results based on two forecatingg erorr metrics, normalized mean squared error (NMSE) and root mean square error (RMSE) shows that proposed hybrid model provides better multi-step-ahead forecasts whilst addressing the issue of long term dependencies. © 2015 IEEE.