Improving Time Series Forecasting Performance by Fuzzy Decision Fusion

Time series data can be collected in many domains including econometric, signal processing, weather forecasting, and earthquake prediction. Accurate prediction of time series prices is essential for investors, meteorologist, or statisticians. Forecasting of the financial time series has intrinsic complexity due to uncertainties of factors that affect it. In this study, better forecasting of the financial stock market time series movements is targeted using the closing prices of the stock market. In this work, our objective is to implement a set of well-known financial time series forecasting models such as Autoregressive-Integrating-Moving-Average (ARIMA), Exponential Smoothing, Support Vector Regression (SVR), Long-Short Time Memory (LSTM), and to merge the forecasted decision by using fuzzy knowledge-based decision system. The difference of this thesis compared to the previous works is mainly the expert-decided membership functions instead of clustering in building the fuzzy rule base. An experimental demonstration has been carried out on the S&P 500 index using the closing prices of this Index. The results shows that the fuzzy decision fusion procedure gives lower cumulative absolute prediction error than cumulative error of forecasts of each individual model. Keywords: time series, time series forecasting, fuzzy decision fusion, fuzzy logic system, fuzzy rule generation.