Hybrid Stock Price Forecasting with Stacked LSTM and Multi-Source Feature Fusion

The stock market (SM), noted for its intricacy and volatility, has historically captivated researchers and investors alike who are in pursuit of effective methodologies to forecast price fluctuations and render informed choices. Precise prediction of SM trends is imperative for optimizing returns and alleviating risks.

Conventional methodologies have predominantly depended on technical analysis, which scrutinizes historical price data, trends and trading volumes to project future market behaviour. Nevertheless, these approaches frequently neglect the significant impact of financial news, social media posts and public sentiments, which may lead to quick changes in the market.

Over recent years, sentiment analysis (SA) [1, 2] has surfaced as a formidable instrument for distilling signals pertinent to the market obtained from text-based sources such as market-related news reports, online social platforms and expert analysis. By quantifying investor sentiment, this methodology offers critical context that augments technical indicators.

Compared to using either technique alone, combining SA and TA has demonstrated a great deal of potential in improving predictive accuracy. The SM is a crucial element of the global economy, working as a platform for capital management and an indicator of economic health. Accurate SM prediction is very important for traders, investors, financial experts and analysts as financial markets grow more complex.

Though widely used, traditional stock market analysis techniques like TA and fundamental analysis (FA) frequently miss the subtle patterns and fluctuations found in financial data. Traditional methods used only historical data of SM for future predictions.

In the modern era, deep learning (DL) has emerged as a transformative approach in the field of SM prediction, offering fresh chances to improve predictive accuracy and dependability.

Stock price (SP) forecasting has been attempted through various approaches, mainly categorized into following three core methodologies as shown below in Figure 1.

tu_pian_42.png

Figure 1. Methods for SM analysis

FA assesses the true intrinsic value of a stock by conducting an analysis of the financial health of company from financial statements and profit performance, growth potential, industry position and broader economic conditions.

To find market trends, patterns and price movements that might point to future price changes, TA looks at historical price data, trading volume and market indicators. These trends are frequently analysed using tools like Moving Averages, Bollinger Bands and Momentum Indicators.

SA, a new field in natural language processing (NLP), examines textual data from news headlines, financial blogs and online forums to determine the general sentiment or opinion of investors. It captures the psychology of the market, which frequently affects price direction without reference to fundamental value. Both FA and TA have drawbacks, such as their incapacity to identify noise in data and to adjust to the quickly changing market conditions.

Stock market forecasting has been transformed by the advent of DL algorithms. Large-scale datasets can contain intricate, non-linear patterns that can be captured by DL models, particularly those that use neural networks. They are particularly beneficial for forecasting time series problems, which forecasts future values based on historical data, because of this capability.

With the evolution of artificial intelligence [2], machine learning (ML) and DL techniques have become popular in SM prediction because of their capacity to record complex temporal dependencies and represent nonlinear interactions. In time-series forecasting, algorithms like Support Vector Machines (SVM), Random Forests (RF), and more recently, DL models like Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN) have produced promising results.

Because of the erratic and unpredictable behaviour of financial markets, predicting SP trends is inherently difficult. However, these predictions are now more accurate than ever before because of advancements in big data analytics, artificial intelligence and processing power. Achieving high forecasting accuracy is crucial because it directly aids in the development of profitable investment portfolios. While DL techniques have shown promising results in identifying trends in large datasets, no model has consistently generated accurate forecasts of future stock prices.

The following are the primary contributions of this research:

1. A hybrid model was developed by combining technical data from four NIFTY 50 stocks from various industries with compound sentiment scores. Stacked Long Short-Term Memory (SLSTM) architecture was used as predictive model and Mean Absolute Error (MSE), Root Mean Squared Error (RMSE) and coefficient of determinant (R²) metrics were used to assess the model.
2. A hybrid model was developed by integrating polarity and subjectivity sentiment metrics with technical data and its performance was evaluated using the same assessment criteria.

Comprehensive experiments were conducted involving multiple runs and hyperparameter optimization. A benchmark configuration was established—specifying ideal neuron numbers per layer, epochs and batch size—to guide future advancements in predictive model design.

The framework used for extracting and analyzing sentiment data from news articles in the Economic Times for selected companies of NIFTY 50, from January 1, 2024, to December 31, 2024. The companies under consideration include Infosys, Sun Pharma, Tata Steel and HDFC Bank.

Figure 2 illustrates a hybrid framework for SP prediction using a fusion of SA and TDS given to an SLSTM model.

tu_pian_43.png

Figure 2. Flow diagram illustrating stock price prediction process

Financial news was pre-processed and analyzed for SA to derive numerical features such as the average compound score and average polarity-subjectivity score. Simultaneously, technical stock data was pre-processed and key features such as open, high, low, close, and volume (O, H, L, C, V) were selected. These sentiment and technical features were fused and fed into an SLSTM model for training and prediction. The proposed model’s forecasting performance was assessed using MAE, RMSE and R² for robust evaluation.

3.1 Data acquisition

The study incorporates two independent data sets. The primary dataset consists of raw stock data from the National Stock Exchange (NSE) for selected stocks. The stock data included parameters such as date, time, opening price, closing price, highest price, lowest price and trading volume, recorded at specific intervals.

The second dataset comprised news articles obtained from the Economic Times.

3.1.1 Financial data

Data collected from NSE India for the time period of 1st January 2024 to 31st December 2024 [36]. Table 1 shows sample data set for TDS.

Table 1. Sample dataset for TDS

3.1.2 Sentiment data

Sentiment data for Infosys, Sun Pharma, Tata Steel and HDFC Bank were collected from the Economic Times news during the same time frame [37].

The Economic Times provides more reliable and domain-specific finance news data compared to Google News, specifically in relation to financial analysis.

3.2 Data preprocessing and feature extraction

The proposed framework presents an automated workflow for conducting SA on finance related news of specific stocks using web scraping and NLP techniques as shown below in Figure 3.

tu_pian_44.png

Figure 3. Block diagram for web scraping and SA score

The process started by selecting a target stock and defining a specific date range to retrieve the relevant news data.

The sentiment analysis process was carried out in Python 3.10, utilizing the Natural Language Toolkit (NLTK v3.9) and TextBlob v0.17.1 libraries. The VADER Sentiment Intensity Analyzer operated with its standard lexicon thresholds (positive ≥ 0.05, negative ≤ –0.05 and neutral = 0). Polarity and subjectivity measures were obtained through TextBlob’s built-in functions, employing default tokenization settings without additional filters. All experimental procedures were executed in Jupyter Notebook version 6.5.4, with model implementation carried out using TensorFlow 2.14 and Keras 3.0. A fixed random seed value of 42 was applied across all tests to maintain consistency and reproducibility of results.

The HTML data was subsequently processed using libraries such as BeautifulSoup to retrieve key components, such as publication dates and news headlines.

This information was structured and stored in a Pandas DataFrame and subsequently saved as a CSV file for further analysis. Table 2 shows the sample dataset for SA.

Table 2. Sample dataset for SA

In the next phase, the saved CSV file containing stock-related news was loaded for SA. Before analysis, the text was pre-processed by removing noise such as punctuation and stop words, converting the text to lowercase and standardizing the format and then SA was carried out by different methods.

In SA the term compound was used to denote the compound score, a normalized value that reflected the aggregate sentiment of a text. This score ranged from -1 (indicating the most negative sentiment) to +1 (indicating the most positive sentiment). It was calculated based on a blend of positive, negative and neutral sentiment scores generated by tools such as VADER.

Let C_n be the compound score for the nth sentence or text. N be the total number of sentiment-scored sentences.

Average Compound Score $=\frac{1}{n} \sum_{n=1}^N C_n$         (1)

Table 3 shows the average compound score of SA. In SA, two essential metrics, the polarity score and the subjectivity score, were often used to figure out how people felt about a piece of writing.

Table 3. Average compound score

The polarity score, which typically ranged from -1 to 1, indicated which sentiment was most prevalent in the text. Scores near -1 denoted extremely pessimistic sentiment where as scores near +1 denoted intense optimism and scores near 0 denoted neutrality.

Tools such as TextBlob were used to evaluate a sentence's or document's emotional tone. On the other hand, the subjectivity score, which ranges from 0 to 1, shows how much the content is influenced by personal opinions rather than objective facts. A score close to 0 indicates an objective or factual tone, whereas a score close to 1 suggests that the text expresses biases, personal opinions or feelings.

Let P_n represent the polarity score of the nᵗʰ sentence or text, Sn denote its subjectivity score and N indicate the total number of sentences for which sentiment scores have been computed.

Average Polarity Score $=\frac{1}{n} \sum_{n=1}^N P_n$          (2)

Average Subjectivity Score $=\frac{1}{n} \sum_{n=1}^N S_n$           (3)

Table 4 shows the calculated average polarity and average subjectivity score of SA.

Table 4. Average polarity and average subjectivity score

In the final step, the average compound score as well as average polarity and average subjectivity score for the news articles of a specific stock for the specific time frame were calculated by combining all sentiment values.

The calculated sentiment scores were saved in separate CSV files (sentiment_stock.csv, for instance), which can subsequently be used for financial prediction or quantitative analysis tasks.

This flexible and scalable framework enabled a data-driven interpretation of market sentiment, helping with tasks like algorithmic trading, investor behavior analysis and risk evaluation by turning unstructured news content into organized and useful insights.

A hybrid model was developed for SP prediction by combining SA and TDS within a Stacked LSTM framework. Financial news was processed and analyzed for sentiment to generate quantitative features, including the average compound score and average polarity-subjectivity score. The following evaluation methods were utilized in SM prediction:

5.1 MAE

It quantifies the average magnitude of absolute differences between predicted outputs and actual observations, without considering the direction of the errors.

MAE is calculated as:

$MAE=\frac{1}{n}\underset{i=1}{\overset{n}{\mathop \sum }}\,\left| {{y}_{is}}-\widehat{{{y}_{is}}} \right|$       (17)

where, n denotes the number of observations.

${{y}_{\text{is}}}$ denotes the variable's actual value at the ith observation, and $\hat{\mathrm{y}}_{\mathrm{is}}$ denotes the expected value of a variable at the ith observation.

5.2 MSE

It represents the average of the squared differences between the observed and predicted values.

MSE is calculated as:

$M S E=\frac{1}{n} \sum_{i=0}^n\left(y_{i s}-\widehat{y_{l s}}\right)^2$        (18)

5.3 RMSE

It is an often-used metric in SM forecasting because of its ability to quantify prediction accuracy, sensitivity to errors, interpretability, comparability and suitability for optimization.

RMSE is calculated as:

$\text{RMSE}=\sqrt{\text{MSE}}$         (19)

5.4 R²

R² is an indicator of the accuracy of predictions in relation to actual data.

R² is calculated as:

${{R}^{2}}=1-\frac{\sum {{\left( {{y}_{is}}-{{{\hat{y}}}_{is}} \right)}^{2}}}{\sum {{\left( {{y}_{is}}-{{{\bar{y}}}_{is}} \right)}^{2}}}$       (20)

Table 5 shows the performance of SLSTM model when integrating compound score with technical data. The best setup for the SLSTM model, which combines stock technical data and compound sentiment scores, is 100 neurons, 100 EP and a batch size of 16, which produces the highest prediction accuracy for all stocks.

Table 5. Performance of stacked LSTM model integrating compound score with technical data

With the lowest MAE values—TATA STEEL (1.72), HDFC BANK (17.64) and INFOSYS (18.34)—this configuration clearly improves forecasting accuracy. INFOSYS (23.85), TATA STEEL (2.16) and HDFC BANK (26.49) all have lower RMSE values, which suggests improved model stability and fewer prediction errors. The R-Squared (R²) values, which reach 0.97 for TATA STEEL and SUNPHARMA and 0.96 for INFOSYS, confirm a robust correlation between expected and actual stock performance.

According to the MAE comparison, prediction errors were greatly reduced by combining polarity and subjectivity features, by as much as 16.3% for Sun Pharma and 2.9% for Tata Steel.

Overall, using more neurons and epochs, training the model with a smaller batch size, and combining sentiment scores with technical indicators greatly improves the model's capacity to comprehend market trend. This results in predictions of stock prices that are more accurate and consistent.

Table 6 shows the performance of SLSTM model when integrating average polarity and subjectivity score with technical data. Using 100 neurons, a batch size of 16, and 100 epochs yields the highest accuracy and consistency in predictions for all examined stocks, according to the evaluation of the SLSTM model, which integrates technical stock indicators with objectivity–subjectivity sentiment measures.

Table 6. Performance of stacked LSTM model integrating polarity–subjectivity score with technical data

With the lowest MAE values, this configuration clearly improves prediction accuracy, especially for TATA STEEL (1.67), INFOSYS (17.04) and HDFC BANK (17.25). Additionally, there are smaller differences between the expected and actual prices, as indicated by the lower RMSE values for TATA STEEL (2.08) and INFOSYS (23.07). Strong agreement between model predictions and actual market data is demonstrated by the R² values, which are still high at 0.98 for INFOSYS and SUNPHARMA and above 0.95 for the majority of other stocks.

An average drop of roughly 10% is seen in the RMSE comparison, suggesting increased prediction accuracy. Overall, the outcomes demonstrate that the model works better when more neurons and epochs are used with a smaller batch size. The hybrid SLSTM model produces accurate and dependable stock price predictions by combining sentiment-based features with technical stock data to better understand market behavior.

Table 7 demonstrates that the SLSTM model outperformed the compound sentiment score when trained with polarity–subjectivity sentiment scores.

Table 7. Comparative study of compound score and polarity-subjectivity scores in a stacked LSTM model

The confidence intervals of 95% for R² were computed using bootstrapping (n = 1000) in order to evaluate reliability. With skewness less than 0.2 and kurtosis less than 3, the error distribution plots displayed nearly normal behavior, indicating steady predictions free from systematic bias.

Using polarity–subjectivity sentiment scores consistently resulted in lower error rates, as shown in Figure 5. As an illustration of improved forecasting accuracy, the MAE for TATA STEEL fell from 1.72 to 1.67 and for SUNPHARMA it fell from 20.31 to 17.00.

tu_pian_46.png

Figure 5. Performance comparison of MAE for SLSTM model for compound and polarity subjectivity score

Lower RMSE values, which indicate smaller prediction errors, were obtained when polarity–subjectivity sentiment scores were used, as shown in Figure 6. For example, SUN PHARMA's RMSE dropped from 25.10 to 21.99, and TATA STEEL's RMSE decreased from 2.16 (with compound scores) to 2.08, indicating less prediction variability.

tu_pian_47.png

Figure 6. Performance comparison of RMSE for SLSTM model for compound and polarity subjectivity score

The R² coefficients derived from polarity–subjectivity scores were more strongly correlated with actual stock prices, as seen in Figure 7. An average improvement of about 0.02 and better predictive alignment were demonstrated by INFOSYS, which obtained an R² of 0.98 as opposed to 0.96 with compound scores.

tu_pian_1.png

Figure 7. Performance comparison of R² for SLSTM model for compound and polarity subjectivity score

5.5 Key contribution of the proposed methodology

1. Improved sentiment representation: This study includes and assesses polarity–subjectivity sentiment features, which are more successful in raising the accuracy of stock price predictions than previous research that primarily used compound sentiment scores.
2. Cross-sector validation: When evaluated on four different NIFTY 50 stocks from the banking, pharmaceutical, steel, and IT industries, the model's resilience and ability to generalize across a variety of market segments were shown.
3. Modified stacked LSTM architecture: A two-layer stacked LSTM model including two dropout layers followed by one dense layer was used, which included 100 neurons,100 EP and batch size of 16. This model gave better predictive performance.
4. Empirical performance gains: Comparisons reveal that polarity-subjectivity features, with MAE as low as 1.67, RMSE as low as 2.08 and R² above 0.96, clearly outperform compound scores in capturing investor sentiment, demonstrating the framework's dependability.

Finally, the integration of technical indicators and polarity-subjectivity sentiment metrics improves the predictive accuracy of the Stacked LSTM model, reduces forecasting errors and aligns actual stock price movements more closely than the use of compound sentiment scores alone. Table 8 shows the comparison of the suggested methodology's performance with cutting-edge techniques

Table 8. Performance comparison of proposed methodology with state-of-the-art methods

The first author carried out all aspects of the research, including conceptualization, data collection, signal processing, model development and manuscript writing, under the guidance of the second author. The second author offered critical feedback, technical supervision and overall direction throughout the study.