Forecasting time series of the market indicators based on a nonlinear autoregressive neural network

The modern practice of economic research relies heavily on mathematical models that make it possible to reveal hidden regularities in statistical data and make forecasts on their basis. Linear models based on vector autoregression (VAR) are the most common. However, the relationship between time series in the economy is often difficult to identify, so non-linear autoregressive (NAR) models show more reliable results. Artificial neural networks (ANNs) are usually used for implementation of these models, but ANNs do not provide the possibility of estimating the forecast in the form of mathematical expectation and a standard deviation. Therefore, the model proposed in the article combines two blocks: VAR and NAR. NAR is used to construct a prediction for a given number of points, and VAR for estimating the forecast in the form of a mathematical expectation and a standard deviation. The evaluation of the model was carried out on the daily data: exchange rate USD / RUB and “Brent” oil from 1.01.2016 to 1.03.2017. The average accuracy of forecasting the trend for the dollar was 54.9%, for the oil prices it was 54.0%. In this case, the relative error in predicting the dollar rate was from 1.09% (for the first point) to 2.01% (for the tenth point); the relative error in forecasting oil prices was from 1.28% (for the first point) to 4.58 % (for the tenth point). Thus, the model showed accurate results when predicting dynamic series and can be used to solve other forecasting problems. In particular, it is expedient to use the model as one of the factors when making investment decisions. In addition, the evaluation of forecasts is done on the basis of testing the NAR block of historical data and on the basis of VAR block forecast in the form of mathematical expectation and standard deviation.

Hali tarjima qilinmagan