Neural networks have been gaining wide popularity in time series forecasting including the prediction of traffic loads. However, there are still some unsolved problems in this filed. The inability to produce and replicate a stable solution, a high probability of overtraining and the absence of strictly defined criteria for selecting a final forecasting model are the most substantial of them. An advanced algorithm introduced in the paper addresses some of these problems and specifies a procedure aimed at facilitating the search of the optimal solution and increasing the reliability of produced forecasts.