Thesis

Leveraging machine learning for financial forecasting : a dual approach for meme stock price and GDP prediction

Downloadable Content

Download PDF
Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2025
Thesis identifier
  • T17333
Person Identifier (Local)
  • 202260979
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • This dissertation explores the adoption of machine learning in financial analysis through a systematic investigation of stock price prediction and GDP forecasting. The first study examines the relationship between online sentiment and meme stock price movements. This is done by conducting sentiment analysis on Reddit’s WallStreetBets discussion thread and identifying daily popular stock tickers and their sentiment scores. The study then uses several recurrent neural networks, including variations of Long Short-Term Memory (LSTM) models such as single-layered LSTM, regular stacked LSTM, bidirectional LSTM model and Gated Recurrent Unit (GRU) model. These models were trained on five years of historical and technical data, highlighting the impact of online sentiment on Meme stock price fluctuations and the potential of AI-driven models to capture these dynamics. Models were tested for real-time applications for three consecutive days. Results demonstrated that the single-layered LSTM model outperformed other models with low error rates. For example, NVDA with average RMSE: 4.64, MAE: 3.38, MAPE: 0.035, and similar performance observed for ASTS with average RMSE 2.03, MAE: 0.92, MAPE 0.091 and LUNR (RMSE: 0.41, MAE 0.28, MAPE 0.066). However, by the third day regular stacked LSTM model slightly outperformed for NVDA, while single-layered LSTM dominated with better predictive power for other stocks (SMCI – RMSE 112.726, MAE: 86.946, MAPE: 0.111; AI – RMSE: 1.6, MAE: 1.197, MAPE: 0.036). The second study extends the use of AI in macroeconomic forecasting, focusing on the prediction of the GDP of the United Kingdom (UK) using vital macroeconomic variables, including energy prices, unemployment rate, inflation, net migration and Real Effective Exchange Rate (REER) from 1990-2018. For the prediction, several machine learning models, such as Support Vector Regression (SVR), Random Forest (RF) and Gradient Boosting Machines (GBM), were implemented and compared together with Shapley Additive exPlanantions (SHAP). The models were assessed using evaluation metrics such as Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE) and R2 score. The findings underscored the significant role of macroeconomic variables in economic forecasting and illustrated the potential of AI-driven models to provide valuable insight into financial markets and economic indicators. Among them scaled SVR model achieved best performance with RMSE: 83,492. 048, MAE: 77, 219.274, MAPE: 4.2% and R2 score of 0.042. Together, these studies demonstrate the adoptability and potential of machine learning in addressing complex financial and economic prediction tasks and underline practical implications. The integration of sentiment analysis for stock price prediction and macroeconomic modelling for GDP forecasting showcases machine learning’s ability to handle diverse data types, from unstructured textual data in online platforms to structured economic indicators. By combining these approaches, the research highlights how AI can uncover hidden patterns and relationships that traditional financial models might overlook, providing a more nuanced understanding of market behaviors and economic trends.
Advisor / supervisor
  • Fernando, Anil
Resource Type
DOI
Date Created
  • 2024

Relations

Items

Thumbnail Title Date Uploaded Visibility Actions
file details PDF of thesis T17333 2025-06-24 Public Download