Evaluation of Deep Learning-Based Long ShortTerm Memory Architecture Variants for Solana Cryptocurrency Price Prediction (SOL/USD)

Abstract

Cryptocurrency markets are highly volatile, posing significant challenges for accurate price prediction. Solana (SOL), one of the largest cryptocurrencies by market capitalization, experiences sharp fluctuations that limit the effectiveness of traditional linear models such as Autoregressive Integrated Moving Average (ARIMA) and Generalized Autoregressive Conditional Heteroskedasticity (GARCH). To address this, the present study evaluates twelve Long ShortTerm Memory (LSTM) architecture variants, including vanilla LSTM, stacked LSTM, bidirectional LSTM, and hybrid models with attention and dropout mechanisms, for forecasting daily closing prices of the SOL/USD pair. Five years of OHLCV data from Coinbase were used, with an $80 / 10 / 10$ split for training, validation, and testing. Performance was assessed using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and $\mathbf{R}^{\mathbf{2}}$. Results show that the LSTM-Base model (60-day window, 1-day horizon) achieved the best accuracy (MAE $=5.38$, RMSE $=6.73$, MAPE $=3.27 \%, \mathbf{R}^{\mathbf{2}}=\mathbf{0. 8 2}$), while LSTM-LB120-H7 (120-day window, 7-day horizon) performed poorly (MAE $=14.09$, RMSE $=17.44, \mathbf{R}^{2}=-0.34$). These findings highlight that simpler LSTM configurations are more effective for volatile crypto assets, offering both academic insight and practical benchmarks for traders, analysts, and policymakers.