Stock Growth Forecast Through Sentiment Analysis on Social Media and News with Statistical Factors

Abstract

The research proposes a comprehensive approach for predicting stock prices over a 30-business-day horizon with an accuracy of 98–99%. By incorporating multiple factors influencing stock prices, the methodology ensures reliable and well-rounded predictions. The base model employs a Univariate Time Series Prediction using a 1D Convolutional Neural Network (CNN-1D), which effectively captures seasonal patterns and general trends. While CNN-1D excels in identifying seasonality, it occasionally struggles with trends. To address this, the research integrates a Kalman Forecaster, which is highly reliable for trend prediction. The combined output of these models provides predictions encompassing both seasonality and trend dynamics, albeit with a slight reduction in accuracy compared to CNN-1D alone. The study further utilizes a Long Short-Term Memory (LSTM) network to predict market indices such as Sensex and Nifty, which, along with CNN-1D and Kalman predictions, inform a Random Forest regressor model trained over 320 days of a 365-day dataset. Predictions for the final 45 days are refined using mean error calculations over the first 15 days, which are then corrected using sentiment scores derived from news analysis powered by a large language model which in this case Gemini Flash Model 1.5. This integrated approach achieves a minimal error margin of 1–3%, successfully combining seasonality, trend, and external market factors for accurate stock price forecasting.