Learning Nonlinear Temporal Patterns in Ethereum Prices Via LSTM Networks

Abstract

A Long Short-Term Memory (LSTM) neural network trained on hourly ETH/USDT market data from the Binance exchange is used in this study to examine short-term Ethereum price behavior. The proposed model emphasizes learning temporal dependencies and momentum-driven structures rather than relying on conventional linear forecasting assumptions, acknowledging the highly nonlinear and noise-dominated nature of cryptocurrency markets. The daily high price of Ethereum is selected as the target variable in the forecasting task, which is defined as a univariate regression problem. To ensure realistic predictive assessment, model performance is evaluated using a strictly out-of-sample testing methodology. Empirical findings demonstrate that the LSTM model achieves a strong statistical fit despite significant market volatility. The obtained results—RMSE of 127.33, MAE of 98.76, MSE of 16,213.76, MAPE of 2.73%, and an R² of 0.96—indicate that a substantial portion of short-term price volatility is effectively captured by the nonlinear architecture. Even in a noise-dominated market, the low MAPE and high coefficient of determination suggest robust predictive alignment. Forecasts over the next five days reveal a recurring short-term directional pattern accompanied by widening prediction intervals, which reflect increasing uncertainty as the forecast horizon extends. This pattern underscores the intrinsic difficulty of achieving accurate price-level forecasts in highly volatile cryptocurrency markets. Overall, when applied to short-term cryptocurrency price dynamics, the results indicate that LSTM models are well-suited for capturing trend persistence and regime-related signals, affirming their usefulness as risk-aware decision-support tools rather than deterministic forecasting systems.