A seasonal-trend decomposition-based dendritic neuron model for financial time series prediction

Financial time series prediction is a hot topic in machine learning field, but existing works barely catch the point of such data. In this study, we employ the most suitable preprocessing technology, machine learning model, and training algorithm to construct a novel seasonal-trend decomposition-based dendritic neuron model (STLDNM) to tackle this issue. The model's unique part is to use the seasonal-trend decomposition based on loess (STL) as preprocessing technology. Particularly, the STL can extract seasonal and trend features from the original data, so that a simple polynomial fitting method can be used to handle these sub-series. Next, the remained complex residual component is predicted by an anti-overfitting dendritic neuron model (DNM) trained by an efficient back-propagation algorithm. Finally, the processed components are added up to obtain the predicting result. sixteen real-world stock market indices are used to test STLDNM. The experimental results show that it can perform significantly better than other previous convinced models under different assessment criteria. This model successfully reveals the internal feature of financial data and certainly improves the predicting accuracy due to the rightful methodology selection. Therefore, the newly designed STLDNM not only has high potentials for practical applications in the financial aspect but also provides novel inspirations for complex time series prediction problem researchers.