MULTIVARIATE TIME SERIES FORECASTING ON FINANCIAL DOMAIN USING SPECTRAL TEMPORAL GRAPH NEURAL NETWORK

• Main Author: Orizadi, Haris
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/63734
• Institution: Institut Teknologi Bandung

Multivariate time series prediction method utilizes information from various time series to generate predictions. Financial domain time series data has characteristics and forms that are influenced by the structure of financial markets and various other attributes. Based on these two statements, multivariate time series prediction was chosen as the method of predicting multivariate time series data in the financial domain by including other attributes that affect price movements of financial data into the prediction algorithm. Financial time series data also influenced by past value of these series. So that by combining prediction methods that can utilize advantages of interrelationships between variables and also recognize temporal patterns in each time series, it is expected to produce more accurate predictions. The characteristics of the financial domain time series data are appropriate with StemGNN architecture (Cao et al., 2021) which integrates graph convolutions in the spectral domain, and temporal convolutions in the temporal domain using Discrete Fourier Transform (DFT). So that in this study an experiment of this method will be carried out in the financial domain. Latent correlation layer (LCL) of StemGNN is the sub-module with the smallest performance contribution, so a trial will be carried out to improve the performance of StemGNN using graph structure formation module from the Multivariate Time Series Graph Neural Network (MTGNN), namely Node Embedding (NE), and comparing it with the graph structure. using the Granger Causality Test (GCT). From the experiments carried out, NE produces better predictive performance than LCL and GCT with an increase in average performance for the rice, commodity, currency datasets of 7.93%, 9, respectively). 73%, and 16.06%.The performance of NE is better when compared to the LCL which uses GRU as the basis for constructing graph structures.