PPS-NET: A NEURAL NETWORK ARCHITECTURE BASED ON POLYNOMIAL PREDICTION SEQUENCES FOR FINANCIAL TIME SERIES FORECASTING

Abstract

The paper considers the problem of forecasting financial time series in conditions of limited data volume and high sensitivity of stock market dynamics to local fluctuations. The neural network architecture PPS-Net (Polynomial Prediction Sequence Network) is proposed, in which a sequence of polynomial forecasts is used as a specialized forecast layer. For each sliding window of the time series, a set of polynomial forecasts of orders from zero to ninth and a sequence of averaged forecast values ​​are formed, among which the final forecast is selected. To reduce the influence of the absolute price level on the selector's operation, structural features are introduced that characterize the relative deviations of the sequence elements from the base forecast. The weight matrix of the selector layer was determined using the weighted regularized least squares method. Weighting of training examples was used to compensate for class imbalance, and the regularization parameter was selected on the validation sample according to the minimum value of the mean absolute percentage error MAPE. To limit the impact of anomalous forecast values, a local admissibility filter was added, which replaces the selected candidate with the base forecast in case of excessive deviation from the local scale of the series change. Experimental verification was performed in Python on intraday stock exchange data of Netflix shares with the ticker NFLX. 780 values of the Close parameter were used for the last 60 trading days with a 30-minute interval. Based on a sliding window of 10 consecutive values, 770 forecast examples were formed, of which 577 were used for training and 193 for testing. The proposed approach was compared with the Δ-method and the DBSCAN PPS clustering method. Compared with the Δ-method, the proposed approach provided a reduction in the mean relative error (MAPE) by 18.47%, and with DBSCAN – by 46.40%. On the last test day, the PPS-Net method reduced MAPE by 11.95% compared to the Δ-method, respectively, for DBSCAN the MAPE reduction is 49.46%. The scientific novelty lies in the integration of a sequence of polynomial forecasts into a neural network selector architecture, the use of PPS structural features and regularized weighting for adaptive selection of the final forecast. The practical significance of the results lies in increasing the accuracy of forecasting intraday financial time series under small sample conditions.