Permuted Temporal Kolmogorov-Arnold Networks for Stock Price Forecasting Using Generative Aspect-Based Sentiment Analysis

Stock prices are experiencing fluctuation daily. While stock price predictions typically rely on historical transaction data, other factors, such as news sentiment, also play an indirect role in influencing these changes. News sentiment, typically expressed as a qualitative sentiment label, cannot be directly used as an indicator in stock price analysis, as it differs from the quantitative stock transaction. The vital issue in stock price forecasting using news data is the quantification process from sentiment label to score. This research proposed generative Aspect-Based Sentiment Analysis (ABSA) to produce an aspect-sentiment quadruplet: aspect category, aspect term, sentiment polarity, and opinion term. The aspect-sentiment quadruplet produces daily ABSA sentiment scores using the Loughran and McDonald Sentiment Lexicon (LMSL) to handle out-of-vocabulary. The stock transaction history and daily ABSA sentiment score are used to unify stock price forecasting using permuted Temporal Kolmogorov-Arnold Network (pTKAN), which is a rearrangement of the position dimension for isolating the sequence of time series. The Movement-Weighted Regression Error (MWRE) evaluation method is proposed to measure the performance of unified stock price forecasting with representation movement direction error and regression error. The experimental results show that the daily ABSA sentiment score positively influences the performance of unified stock price forecasting using the Temporal Kolmogorov-Arnold Network (TKAN). The pTKAN architecture best performed in 25 of 27 stock issuers among 19 architectures, which includes traditional-, machine learning-, and deep learning-based architectures tested on the stock transaction dataset.