Forecasting news sentiment-oriented stock market based on sequential transfer learning approach

Bea, Khean Thye (2023) Forecasting news sentiment-oriented stock market based on sequential transfer learning approach. Master dissertation/thesis, UTAR.

Preview

PDF Download (1446Kb) | Preview

Abstract

Forecasting stock movements is a significant and challenging task due to the dynamic and high variability of market attributes. The massively available online text information holds the key to reveal the unexplained variability and facilitate the forecasting accuracy. However, text data is largely unstructured and exists in the form of natural language, necessitating an alternative way to process and capture the insight. This study tapped into the potential of cutting-edge Neural Networks and Natural Language Processing (NLP) techniques, with a specific emphasis on the sentiment-based approach for financial forecasting. Sequential transfer learning is a recent advancement in neural networks applied to Natural Language Processing (NLP). It adopts the “pre-train then fine-tune” paradigm to leverage existing knowledge to enhance the performance of different downstream NLP tasks and achieve state-of-the-art results. Our objective is to assess the performance of various pre-trained models such as BERT, FinBERT, and SKEP, in the context of sentiment-based financial forecasting. This study is novel in its application of these models to the scenario of algorithmic trading, offering a fresh perspective in this research area. Specifically, these pre-trained models served as the sentiment analyzers to transform the financial news titles into sentiment features. These are then combined with a range of technical indicators and used as input to the (MRM-LSTM) predictive model. The KLCI market index is primary prediction focus in this study to demonstrate the effectiveness of the proposed approach. The findings indicate that the inclusion of sentiment features extracted from financial news via these pre-trained models results in a substantial improvement in the accuracy of KLCI index price movement predictions. These results are statistically significant in achieving the highest average point return of 336.10 with F1 score of 53.59, underlining the validity of our approach. In summary, this study finds that sequential transfer learning is effective in extracting sentiment features from financial news and provides superior financial predictive performances when use in conjunction with other market data.

Actions (login required)