Solar Shocks and Predictive Ceilings: A Comparative Benchmarking of Machine and Deep Learning Architectures in the Egyptian Solar Energy Stock Price Market

The rapid expansion of the Egyptian renewable energy sector has created a critical need for high-precision financial forecasting tools to guide institutional investment and policy stability. This study presents a comparative benchmarking of five computational paradigms, Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), Multi-Layer Perceptron (MLP), Random Forest, and XGBoost to predict the stock price evolution of six major solar energy entities: ACWA Power, JA Solar, Jinko Solar, Longi, TAQA, and Trina. Utilizing a daily historical dataset spanning from 2019 to early 2026, we transformed raw price sequences into stationary log returns to mitigate the impact of market noise and heteroscedasticity. Our empirical findings reveal a distinct performance hierarchy based on the asset's volatility regime. The LSTM model emerged as the most robust architecture for stable, trend-driven entities like JA Solar and Trina, achieving R^2 scores above 0.95 and MAPE values below 2%. Conversely, all models exhibited a predictive ceiling when applied to ACWA Power, where sudden "step-shifts" and regime changes led to negative R^2 values, highlighting the limitations of pure historical price action in capturing fundamental market shocks. By performing an inverse log transformation to reconstruct actual price levels, this research provides a practical framework for identifying which algorithmic structures are best suited for different market conditions. These results offer actionable insights for financial analysts navigating the complexities of emerging solar markets, suggesting that while deep learning effectively tracks cyclical growth, hybrid models incorporating fundamental sentiment are necessary for shock-driven assets. JEL Classification: C45, C53, C58, G17, Q42, Q47