Benchmarking Liquid Neural Networks Against LSTM and GRU for Time-Series Prediction

This study presents a comprehensive benchmark of Liquid Neural Networks (LNNs) against traditional Recurrent Neural Networks (RNNs), specifically Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) architectures, for time-series prediction tasks. LNNs are biologically motivated architectures featuring dynamic and adaptable behaviour due to continuous-time dynamics. The research evaluates these models on three datasets: historical Apple Inc. stock price data (AAPL), a synthetic sine wave series, and synthetic temperature data, based on predictive performance (Root Mean Squared Error — RMSE, Mean Absolute Error — MAE), training consistency, and computation speed. Results show that GRU outperforms on stock price prediction (RMSE = 5.06; MAE = 4.35), while LNN achieves superior accuracy on synthetic datasets (sine wave: RMSE = 0.101, MAE = 0.079; temperature: RMSE = 0.520, MAE = 0.424). LNNs exhibit the fastest inference speeds (0.000281–0.000299 s/sample) while maintaining training times comparable to LSTM.