A Deep Learning Model For Financial Market Forecasting: FEPA Model

The current global economies and financial markets come into integration. Financial markets continue to show a lot of complex phenomena that the classical finance theory is unable to explain. Financial markets are no longer an ideal market for efficient market hypothesis(EMH). Financial markets show the characteristics of high intelligence, strong correlation and tight coupling which make it a complicated nonlinear dynamic system. To describe such a complex nonlinear dynamic systems, to reveal the inherent laws of financial time series, and to show its evolutionary mechanisms before the world, to facilitate people to prevent financial risks, to supervise and regulate the market, this research on financial time series forecasting, is no doubt, of great practical significance and theoretical value.Recently, deep learning has become the most powerful technique of artificial intelligence for developing intelligent system models. AlphaGo of Google DeepMind has just beaten the world champion of Go, the most difficult of complete-information games. However, it is known that financial markets are far more complex than all kinds of board games. This PhD thesis documents an investigation into one of the three paradigms of deep learning, Deep Decomposition Synthesis Neural Networks(DDSNN) for financial market forecasting. In this work, DDSNN takes the actual form of FEPA Model – FTS-EMD+PCA+ANN. More details are given below.Huang et al(1998) from NASA proposed a new signal analysis method: Empirical Mode Decomposition(EMD). As a better followup to Fourier transform and wavelet transform, EMD method does not set basis functions in advance and decomposes the time series in accordance with its own scale characteristics. In Fourier transform, harmonic basis functions are used. In wavelet transform, particular wavelet basis functions are also needed in advance. EMD method further improves the local characteristics of time series based on Fourier transform and wavelet transform. It is a more advanced data mining preprocessing algorithm. As the EMD method has better characteristics to process the data, any type of signal can be decomposed by EMD method. Therefore, it has a very clear advantage for signal processing of the data with nonlinear non-stationary and multi-scale characteristics. After its invention, EMD has been effectively applied to many engineering fields.Firstly, the new combined forecasting model based on EMD decomposition proposed in this thesis is called: FEPA model.(FTS-EMD-PCA-ANN). The model is made up of three components: financial time series special EMD(Financial Time Series Empirical Mode Decomposition, FTS-EMD), principal component analysis(Principal Component Analysis, PCA) and artificial neural networks(Artificial Neural Network, ANN). The FEPA model is designed to forecast nonlinear non-stationary multi-scale complex financial time series. The FEPA model is implemented to predict the stock indices and the foreign exchange rates. The combination forecasting model put forward in this thesis is based on the idea of decomposition and synthesis for forecasting signal reconstruction and integration. The model effectively improves the prediction accuracy of the financial time series. For testing, the CSI 300 stock index and the Australian stock index as well as foreign exchange rates are chosen as target markets and historical data in this thesis. Six forecasting models are used: ARIMA, GARCH, BPNN, EMD-BPNN, WD-BPNN, EMD-LPP-BPNN to compare with the FEPA. The short-term moving tendencies of the close price are predicted respectively by these models. The empirical results show that the EMD-BPNN model has better forecasting performance than ARIMA, GARCH and BPNN models. This shows that the forecasting method of the first decomposition and then combination can improve the prediction accuracy of the nonlinear and non-stationary financial time series. The testing results also show that the new FEPA model proposed in this thesis has even higher prediction precision and reliability than the EMD-BPNN model. This shows that principal component analysis can compress redundant data, shorten training time and improve prediction accuracy.Secondly, most existing models of financial time series analysis only consider to analyze the close price, but the stock index oscillate between the highest price and the lowest price every day. If we only analyze the close price, a lot of useful information will be waisted. This can cause the prediction information not comprehensive and may not meet the requirements of market investors. An interval EMD decomposition is introduced as an algorithm which considers to analyze the highest price and the lowest price simultaneously. The volatility and the trend of the stock index time series can be preferably grasp by analyzing the close price, the highest price and the lowest price of the stock index simultaneously. The test results show that when the interval EMD decomposition algorithm is applied to FEPA model, it can effectively improve the prediction effect of the highest and the lowest price.Thirdly, most of the econometric models in the existing literature are designed for the same time scale of training data. This ignores the multi-scale features of financial time series. Here a multi-scale information fusion model is introduced. The CSI 300 index close price of the daily data and the 15-minutes frequency data are used to form the input to neural networks for traning simultaneously Two benchmark ANN models are also developed using daily data and 15-minutes data respectively. The test results show that the multi-scale neural network model has better prediction performance. Of course, this kind of multi-time scale integration of financial information has a huge research space.Finally, taking into account the financial and economic globalization, the world financial markets are more closely related to each other. Even if to predict the trend of a single market, it also needs to find lead markets which have significant impact on the single market. On the basis of an in-depth study of Copula theory, this paper applies the theory to the asymmetric correlation analysis of the world’s major stock market indexes. Asymmetric correlations among the selected stock indexes are modeled and detected using the relevant metrics of the Copula function on the logarithmic yield of stock indexes; The detected asymmetric correlations are put together to form a directed acyclic graph. Test is done on historical daily data with the results showing that the Copula correlation coefficients are more informative for finding the influential leading markets for the predefined target market better than the traditional linear correlation coefficients.