Time‐Delayed Spiking Reservoir Computing Enables Efficient Time Series Prediction

Spiking reservoir computing (SRC), a variant of reservoir computing in which the reservoir is realized as a spiking neural network, demonstrates competitive performance on time series prediction tasks. However, the computational cost increases substantially with reservoir size, limiting its scalability. In this work, time‐delayed spiking reservoir computing (TDSRC) is proposed to address this limitation. This method constructs an expanded readout feature vector by concatenating time‐lagged reservoir states, enabling comparable prediction accuracy with smaller reservoirs. TDSRC is evaluated on three benchmark tasks: pulse‐driven function generation, Lorenz system forecasting, and Mackey‐Glass system forecasting. Across all tasks, TDSRC achieves prediction errors comparable to or lower than those of standard SRC with fewer neurons, indicating a trade‐off between the time delays and the reservoir size. Furthermore, the effects of delay selection and the excitation–inhibition ratio of spiking neurons in TDSRC are analyzed, providing practical guidelines for its implementation.