Long-Memory Reservoir Computing for Data-Scarce Dengue Forecasting

Accurate dengue forecasting is hindered by the unique nature of epidemic data: it is often non-stationary, nonlinear, and characterized by long-range temporal dependence (long memory). While classical statistical models like ARFIMA can capture long memory, they struggle with nonlinear dynamics. Conversely, deep learning models like LSTMs can capture nonlinearities but typically require large datasets and often fail to represent true long-memory processes, as their hidden states tend to decay exponentially. This paper addresses the challenge of forecasting dengue incidence in data-scarce environments by developing a reservoir computing framework that balances these competing requirements.

The authors propose a long-memory reservoir computing framework based on Echo State Networks (ESNs). ESNs are computationally efficient because they keep the recurrent reservoir fixed and only train a simple linear readout. To address the "short-memory" limitation of standard ESNs, the authors introduce two variants:

• Fractional ESN (fESN): Incorporates fractional-differencing dynamics directly into the reservoir to encode long-range dependence through polynomial decay.
• Wavelet ESN (wESN): Uses wavelet smoothing to extract stable low-frequency components from the data before feeding them into a memory-aware reservoir.

The authors provide theoretical proofs demonstrating that while standard ESNs generate short-memory processes, their proposed variants successfully generate long-memory processes. They also integrate conformal prediction to provide calibrated uncertainty intervals for their forecasts.

Across nine different dengue incidence datasets, both fESN and wESN consistently outperformed traditional statistical baselines (like ARIMA/ARFIMA) and various deep learning models (RNNs, LSTMs, GRUs). The fESN model generally achieved the best overall performance. The study confirms that the proposed models are particularly effective in data-scarce settings where traditional high-capacity neural networks fail to generalize. Furthermore, the inclusion of conformal prediction allows for robust, distribution-free uncertainty quantification, which is critical for public health decision-making.