Researchers' fESN and wESN models outperform deep learning on data-scarce dengue forecasting

Accurate dengue forecasting is critical for public health planning, but it's notoriously difficult because incidence data is often short, noisy, non-stationary, and exhibits long-range dependence. Traditional methods like ARFIMA handle non-stationarity and persistence but are linear and miss nonlinear patterns. Deep neural networks capture nonlinearity but require large datasets that simply aren't available for many outbreaks. Echo State Networks (ESNs) offer a middle ground—they keep nonlinear recurrence while training only a simple readout—but standard ESNs lack the long-term memory needed for this specific time-series problem.

To bridge this gap, researchers introduce a long-memory reservoir computing framework with two novel variants: Fractional ESN (fESN) and Wavelet ESN (wESN). fESN directly encodes fractional differencing into the reservoir dynamics to model long-range dependence, while wESN uses wavelet smoothing to extract stable low-frequency components before feeding them into a memory-aware reservoir. Crucially, they provide theoretical guarantees that standard ESNs induce short-memory processes, but their new reservoirs produce polynomially decaying dependence consistent with statistical long memory. Across multiple dengue datasets and forecasting horizons, both fESN and wESN consistently outperform statistical and deep learning baselines. When combined with conformal prediction, they also deliver distribution-free, calibrated uncertainty intervals—a practical necessity for public health decision-making.