A Novel Multivariate Bi-LSTM model for Short-Term Equity Price Forecasting

The stock market is a critical component of the financial system, and its significant price fluctuations can have far-reaching effects on the broader economy [1]. Accurate predictions of stock prices are essential for preventing market crashes, enhancing market management, and fostering financial stability. However, forecasting stock prices presents challenges due to the market’s volatility, non-linearity, and the complex nature of financial data. While traditional methods have focused on historical data analysis to predict future trends, recent advancements in machine learning, particularly Long Short-Term Memory (LSTM) networks, have shown promise in improving prediction accuracy by capturing long-term patterns in time-series data [2, 3, 4]. This study utilizes both Univariate and Multivariate LSTM models to forecast short-term stock prices, with a focus on stocks from the National Stock Exchange of India (NSE). Given the high volatility of the stock market, effective prediction mechanisms are crucial for managing risks and making informed investment decisions, especially in India where over 190 million people actively trade stocks [8]. By analyzing historical market data, this study aims to improve resource management and trend forecasting. Hourly data is used to capture daily market fluctuations while considering computational complexity, with the goal of identifying trading opportunities and enhancing short-term price trend analysis. The research focuses on predicting short-term closing prices for stocks from four major sectors—Finance, Energy, Industrial, and Information Technology (IT)—which are significant contributors to India’s economy. The selected stocks are among the top contributors to the NIFTY 100 index, making them key indicators of market performance. Additionally, backtesting is employed to evaluate the effectiveness of trading strategies and predictive models using historical data. This process helps researchers and investors assess the potential success of their approaches before real-world application, thereby reducing financial risk and improving decision-making accuracy.

Stock price prediction has evolved from traditional statistical methods like ARIMA (AutoRegressive Integrated Moving Average) and GARCH (Generalized Autoregressive Conditional Heteroskedasticity) to Machine Learning (ML) techniques and, more recently, deep learning approaches to enhance forecasting accuracy [9]. Statistical methods, foundational in econometrics, capture patterns in historical data but may struggle with non-linear relationships and complex patterns in modern markets. For instance, Menon et al. (2016) successfully used ARIMA and GARCH for bulk price forecasting [10], while Shakhla et al. (2018) applied multiple linear regression with success [11].
ML models, including Naive Bayes(NV),Random Forests (RF) and Support Vector Machines (SVM), are able to overcome these limitations by leveraging their ability to handle large datasets and complex variable interactions [12]. Patel et al.(2015) applied these algorithms to predict stock and stock index movements, demonstrating superior performance compared to traditional statistical approaches. However, despite their advantages, these methods often require extensive hyperparameter tuning, can be computationally intensive, and suffer from overfitting, especially when applied to volatile financial markets.
ANNs were designed to learn intricate relationships through multiple hidden layers and activation functions and used by Guresen et al. for stock market index prediction. These models can handle large volumes of data with high dimensionality and uncover deep, non-linear dependencies that ML techniques might miss[13]. However, a drawback of ANNs is that they cannot deal with temporal dependencies and sequential patterns.
Recurrent Neural Networks (RNNs), overcoming these temporal dependencies, handle sequential data by maintaining a form of memory through feedback connections. This enables them to process sequences more effectively than ANNs. Though RNNs can efficiently handle temporal data[14], they face issues like vanishing gradients, which affect their performance over long sequences.
Long Short-Term Memory (LSTM) networks are a major improvement for time series prediction, addressing the issues faced by RNNs. LSTMs include memory cells capable of retaining information over extended periods, lessening the vanishing gradient problem and allowing accurate capturing of long-term dependencies. Hochreiter and Schmidhuber (1997) first introduced LSTM networks, and emphasized that they could represent intricate temporal patterns[4]. LSTM models possess short-term memory and can catch the longer term effects and predict financial data accurately[15]. LSTMs outperform traditional econometric models in forecasting foreign exchange rates[16]. Additionally, LSTMs can achieve better results than classical ML methods in predicting stock market movements[17]. Taking these factors into consideration, this study focuses on LSTM as the primary architecture.
Unlike traditional LSTMs that process data in a single direction, bidirectional LSTMs analyze sequences in both forward and backward directions [19]. This approach allows them to utilize information from both past and future contexts, offering a more comprehensive understanding of the data. This makes them especially efficient for stocks [20]. Notably, Han et al. implemented a Bidirectional LSTM model, with a mean squared error of 0.00020 [21]. We propose to use technical indicators to further improve upon the model performance.
Incorporation of technical indicators, like moving averages and relative strength index, leads to further improvement in the accuracy of prediction models. Alsubaie et al. in their 2019 study, found that using at least 10 technical indicators will lead to a higher prediction accuracy on stock price data[22].
In this study, we present a Bidirectional Multivariate LSTM model for short term trade prediction, which incorporates the use of technical indicators.

This paper aims to develop a novel technique for forecasting stock price trends, specifically tailored for short-term trades using hourly data from ICICI, NTPC, Ambuja Cement, and Wipro over an eight-year period (2015–2022). We employed Unidirectional and Bidirectional LSTM models for both Univariate and Bivariate data. Three research approaches were explored: the first used a basic univariate forecast model with stock close price data, the second included additional features such as open, high, low, and volume, and the third incorporated 12 features, including technical indicators that closely follow the trends of the close price.
The analysis revealed that the proposed Bidirectional LSTM model outperformed Unidirectional LSTMs, demonstrating a comprehensive understanding of price movements. Furthermore, the inclusion of technical indicators significantly improved the proposed model’s performance across all sectors of the Indian National Stock Exchange, demonstrating it’s robustness and adaptability to diverse market conditions.
The proposed model achieved the highest $R^{2}$ score of 99.6736% on Ambuja Cement, underscoring the value of integrating technical indicators with advanced LSTM architectures. This makes it ideal for predicting stock movements in real time, leading to more profitable and secure investments across all major sectors of the Indian National Stock Exchange.
The proposed model offers potential for future improvements and wider applications. One approach is to extend it for long-term stock price predictions by using longer window sizes to capture broader market trends. Expanding the model to other major stock exchanges, like NYSE, Nikkei, and LSE, could validate its effectiveness across different market conditions. Additionally, incorporating macroeconomic indicators, social media sentiment, and global news could enhance its predictive power, leading to a more comprehensive approach to stock price forecasting.