Umeå universitets logga

Unpredictable and limited information are critical reasons behind the challenges faced in forecasting one-day-ahead stock prices across S&P 500 financial sector. Through this thesis, we propose a structured deep-learning mechanism which combines twenty-days-worth of sliding windows of per-ticker-normalized closing price across nine quarterly financial ratios. The financial ratios include P/B (price-to-book), EPS (earnings per share), P/E (price-to-earnings), EV/EBITDA (Enterprise Value-to-EBITDA), PEG, Beta, ROE (Return on Equity), ROA (Returnon Assets), and Debt-to-Equity. The study proposes three models based on data collected via seventy-three stocks. The goal is to to evaluate whether the integration of quarterly fundamental features with technical price windows improves prediction accuracy across the financial sector of the S&P 500, in comparison torelying solely on technical data. Each model is assessed using standard error metrics along with the share of variance it explains. Hybrid architecture model show cases highly significant performance in comparison to the single LSTM as well as pure CNN primarily due to lowered errors in predictions while it effectively explains a larger number of variability in target sections. The thesis further finds that the effectiveness of the hybrid setup diminishes when small random variations are applied to technical inputs without ensuring adjustments to fundamental data, especially since such disturbances interfere with crucial signals. The research findings further indicate that integrating technical features with fundamental concepts helps generate reliable one-day-ahead predictions in comparison to only focusing on price information.