V. V. Kolmykov

Nowadays the problem of forecasting in insurance business has become actual for the formation of the insurance bonus is affected by the external factors. At first the factors that affect the system are chosen. On the basis of the expert estimation there were chosen 8 factors that have the biggest impact on collecting of the insurance bonus. After formation of the output data the backpropagation was chosen. On the basis of the backpropagation the network learning was carried out and using the one-step forecasting the prognosis over the next period was built. While creating a neural network the attention is paid on the number of connections because it should be that the overall number of connections has to be several times less than the learning sample. Otherwise the neural network will be just "overlearned" what means that it will remember the data having lost the capability of making statistically significant forecasts based on the new data. In this paper there have been chosen networks in which the number of connections was bigger than the learning sample size because there were found 8 factors affecting the formation of insurance bonus. If the number of connections is decreased the number of factors has to be decreased what causes not all of the factors to be taken into consideration. In the experiments there were used neural networks of 8-3-2-1 (8 inputs, 2 hidden layers with 3 neurons in the first one and 2 neurons in the second one and 1 output) and 8-3-1 (8 inputs, 1 hidden layer with 3 neurons and 1 output) types. On the assumption of all the experiments it was found out that the network of 8-3-3-1 type gives the insurance bonus error of 5.1% and the insurance bonus error of the network of 8-3-1 type is 10.9%. Most precisely the tendency of the insurance bonus formation is forecasted by the network of 8-3-2-1.