Stock Price Prediction Accuracy with Hybrid BiLSTM Enhanced Transformer-TCN Model

Abstract

This paper presents an advanced AI-driven system for forecasting short-term stock prices using a hybrid deep learning architecture that integrates Bidirectional Long Short-Term Memory (BiLSTM) networks, Transformers, and Temporal Convolutional Networks (TCN). The proposed model is designed to enhance prediction accuracy by capturing both long-range dependencies and local temporal patterns in financial time series data. Two years of historical stock data are sourced from Yahoo Finance, including Open, High, Low, Close, and Volume values. The data is preprocessed using Min-Max scaling and structured into 60-day sliding windows to train the model on sequential trends. The BiLSTM component captures bidirectional temporal relationships, the Transformer applies attention mechanisms to focus on relevant data patterns, and the TCN extracts high-resolution short-term features through dilated convolutions. In addition to deep learning-based forecasting, the system incorporates Donchian Channels to detect potential trend breakouts and reversals, offering users clear insights into stock momentum. A web-based interface provides real-time visualization of predictions, trends, model accuracy, and candlestick charts with technical indicators. Testing results demonstrate high predictive accuracy ranging between 85% and 95%, validating the robustness of the hybrid model. This work highlights the effectiveness of combining multiple deep learning paradigms in financial forecasting.