An LSTM-Based Framework For Short-Term Solana Price Prediction

Abstract

This study examines short-term Bitcoin price dynamics using a Long Short-Term Memory (LSTM) neural network trained on hourly SOL/USDT data from the Binance exchange. To address the nonlinear and noise-dominated nature of cryptocurrency markets, the model is designed to capture momentum patterns and temporal dependencies rather than rely on linear forecasting assumptions. The prediction task, framed as a univariate regression problem with the daily high price as the target variable, is evaluated using a strictly out-of-sample framework. Empirical results show that despite considerable market volatility, the LSTM model demonstrates strong statistical performance. The model’s RMSE of 10.53, MAE of 8.76, MSE of 110.92, MAPE of 6.09%, and R² of 0.78 indicate that its nonlinear architecture captures a substantial portion of price volatility. In terms of absolute price, the model achieved an RMSE of USD 3,619.74, an MAE of USD 2,989.73, a MAPE of 2.82%, and an R² of 0.90, demonstrating its robustness in both normalized and real-scale assessments. Forecasts suggest that SOL prices are likely to decline from USD 88,425.62 to USD 85,497.88 over the next five days, reflecting a persistent short-term bearish trend with wide prediction intervals of around USD 6,000, indicating substantial uncertainty. These findings imply that LSTM models can effectively extract trend and regime-related information, making them valuable as risk-aware decision-support tools rather than deterministic forecasting systems—even though accurate short-term price-level prediction remains challenging.