A METHOD FOR PREDICTING STOCK PRICES USING BILSTM AND AN ENHANCED TRANSFORMER

The financial industry has, for a very long time, placed a significant emphasis on the optimization of share
holder returns. In order to enhance the precision and dependability of stock price forecasting, this article presents a
novel model known as BiLSTM-MTRAN-TCN. By adding Temporal Convolutional Networks (TCN), the suggested
technique revises the conventional transformer model. This results in the creation of a unique transformer version
(MTRAN-TCN) that is specifically designed for stock market forecasting. This model takes use of the benefits that the
BiLSTM, transformer, and TCN architectures have to offer by combining them with MTRAN-TCN. BiLSTM stands
for "Bidirectional Long Short-Term Memory." In spite of the fact that transformers are particularly effective at
collecting long-range relationships, they are not very good at processing sequential information. BiLSTMs, on the other
hand, are able to capture sequence patterns that are bidirectional, and TCNs improve the model's capacity to generalize
by successfully representing sequence dependencies. The enhanced performance of the transformer as well as the
advantages of integrating BiLSTM were confirmed by employing five index stocks and fourteen equities from the
Shenzhen and Shanghai exchanges. This technique yields much superior outcomes across a variety of stock indexes
when compared to other models previously published in the academic literature. The technique was able to reach the
greatest R2 score in 85.7% of the stock datasets, with a decrease of 24.3% in root mean square error (RMSE) to 93.5%
and an increase of 0.3% in the R2 value to 17.6%. Additionally, the model displayed consistent predicted accuracy
across a variety of time periods without causing any worries regarding the timeliness of those predictions. Based on
these findings, it is clear that the BiLSTM-MTRAN-TCN model provides superior performance in terms of stock price
forecasting. It demonstrates both high accuracy and broad generalization capabilities.