A novel incremental rough set-based pseudo outer-product with ensemble learning (ieRSPOP) neuro-fuzzy system for forecasting volatility

Abstract

The rough set pseudo outer-product fuzzy neural network (RSPOP FNN) is a member of the POPFNN family known for high accuracy and interpretability, and also uses rough set theory to perform attribute and rule reduction. An incrementalensemble RSPOP FNN, named ieRSPOP, is proposed and implemented in this project. This new system aims to further improve the capability of RSPOP by using an incremental learning algorithm. Issues with incremental rough set attribute reduction are also addressed using ensemble learning. ieRSPOP utilizes the compositional rule of inference (CRI) method due to its dominance in the field of approximate reasoning. The performance of ieRSPOP is evaluated through several time series benchmark experiments and stock data and analyzed against existing incremental and non-incremental architectures, and results are promising. The project also attempts to forecast real life stock price volatility. The concepts of econometric models of generalized auto-regressive conditional heteroskedasticity (GARCH) and intraday volatility indicators are used. When combined with GARCH concepts and indicators, the performance of ieRSPOP approaches the accuracy of well-known GARCH models. ieRSPOP also provides the added benefit of generating IF-THEN fuzzy rules which describe the GARCH volatility model.