Keywords
Orthogonal Regularization
Time Series
How to Cite
Abstract
The paper proposes a new method for forecasting foreign exchange time series, which improves accuracy and robustness against noise and non-linear structures present in the data. The main innovation is the integration of an orthogonal transformation layer (e.g., Fourier, Wavelet, PCA-based) with traditional RBF and MLP layers. This hybrid approach ensures optimal data preprocessing, reduces feature correlation, and minimizes the conditional number of the Jacobian matrix. A significant theoretical and practical contribution is the introduction of orthogonal regularization, which directly controls the preservation of the orthogonality of the transformation matrix during learning. Experimental results on three major currency pairs (EUR/USD, GBP/USD, USD/JPY) confirm that the proposed hybrid model significantly reduces the Mean Squared Error (MSE) by 12–15% compared to classical models and accelerates convergence by 20–30%. The paper highlights this model's superiority over LSTM and CNN networks in terms of its robustness and computational efficiency
References
Bishop C. M. Pattern Recognition and Machine Learning. Springer, 2006.
Haykin S. Neural Networks and Learning Machines. Pearson, 2009.
Goodfellow I., Bengio Y., Courville A. Deep Learning. MIT Press, 2016.
Box G., Jenkins G. Time Series Analysis: Forecasting and Control. Wiley, 2015.
Mallat S. A Wavelet Tour of Signal Processing. Academic Press, 2009.
Zhang G. Time series forecasting using a hybrid ARIMA and neural network model. Neurocomputing, 50 (2003).
Hornik K., Stinchcombe M., White H. Multilayer feedforward networks are universal approximators. Neural Networks, 1989.
Poggio T., Girosi F. Networks for approximation and learning. Proceedings of the IEEE, 1990.
Oseledets I. V. Tensor-train decomposition. SIAM Journal on Scientific Computing, 2011.
Kloeden P., Platen E. Numerical Solution of Stochastic Differential Equations. Springer, 1992.
Kingma D., Ba J. Adam: A method for stochastic optimization. ICLR, 2015.