Fourier Graph Convolution Transformer for Financial Multivariate Time Series Forecasting

Financial Multivariate Time Series (Fin-MTS) forecasting is increasingly critical in the financial market. Unlike other Multivariate Time Series (MTS) data, Fin-MTS exhibits particular characteristics, including non-linearity, volatility, and hidden periodicities, which thus introduce great challenges for modelling it well. Existing state-of-the-art models for Fin-MTS forecasting often overlook hidden periodic characteristics, such as credit and monetary policy cycles. More importantly, these models usually show limited capability in well capturing the intra-series and inter-series dynamic information during the modelling process, resulting in significant information loss in quantitative finance modelling and thus limited forecasting performance. To this end, in this paper, we introduce a novel model called Fourier Graph Convolution Transformer (FreTransformer) for Fin-MTS modelling and forecasting. FreTransformer is not only able to well model both the intra- and inter-series dynamic dependencies, but also well capture the important hidden periodicities embedded in Fin-MTS data. FreTransformer first maps the original time domain data into the frequency domain to disclose the hidden periodicities and then employs a novel Fourier Graph Convolution Network to well capture the complex intra- and inter-series dependencies within Fin-MTS. Extensive experiments on real-world US market data across 12 phases demonstrate that our method outperforms current state-of-the-art models. Our source code is publicly available at this repository: https://github.com/AmsonntagChow/FreTransformer