Financial time series forecasting based on momentum-driven graph signal processing

Forecasting is important for social development and industrial production in today's complex and fluctuating economic environment. The nonlinearity and non-stationarity of financial time series (FTS) data make it difficult to achieve accurate prediction. This work proposes a forecasting method for the return of FTS based on the emerging field of graph signal processing (GSP). The proposed method makes forecasting decisions based on the similarity between the current tendency and historical tendencies of FTS data. First, a topological graph is created based on the underlying structural relationship of different historical FTS datasets. Subsequently, spectral clustering is used to select historical datasets that are similar to the current dataset, and then the future values are predicted by weighted averaging the selected historical samples. In addition, a momentum-driven method is introduced to improve the robustness of the forecasting results. Finally, the proposed method is extended to a collaborative forecasting framework, where auxiliary macroeconomic data is introduced to improve the forecasting accuracy. The superiority of the proposed methods is verified by a set of numerical experiments with different stock indices.