Research map / Time series / Causal and generative time series

Reservoir Computing for Short High-Dimensional Time Series: an Application to SARS-CoV-2 Hospitalization Forecast

ICML 2024 · Time series › Causal and generative time series · approach: Robust forecasting and evaluation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Applies reservoir computing combined with genetic-algorithm feature selection (RC-GA) to forecast 14-day SARS-CoV-2 hospitalizations from 409 predictors over 586 days, outperforming RC alone and LSTM, Transformers, Elastic-Net, and XGBoost.

Problem

Forecasting SARS-CoV-2 hospitalizations 14 days ahead in a short, high-dimensional time series setting (409 predictors, 586 daily observations) where conventional models perform poorly.

Approach

Pairs echo-state-style reservoir computing with a genetic algorithm that jointly tunes reservoir hyperparameters and performs feature selection over the high-dimensional input, aggregating many reservoir instances for stable forecasts.

Key results

• RC-GA with 20 hyperparameter reservoirs and monthly update reached MAE 14.70, and RC-GA without monthly update reached MAE 15.27 (14.66 with 2000 units).
• RC-GA without feature selection (MAE 17.59) and with random search instead of GA (MAE 18.45) performed much worse.
• Baselines: XGB-RS 15.45, Enet-RS 15.83, LSTM-PCA 15.74, Transformers-PCA 19.28; RC-GA also outperformed XGB at 7 and 21 days horizons, though both were worse than baseline at 21 days.

Measured effects reported in the paper

Surprising findings

• Both 21-day models performed worse than the naive baseline of persistence at 14 days' current value.
• Monthly hyperparameter updates did not improve most models, and sometimes hurt RC-GA, attributed to GA convergence to a high leaking rate.

Limitations

• Single hospital and single epidemic period, so generalizability is untested.
• Performance is sensitive to the GA hyperparameter search and reservoir count; variance across reservoirs is large (std about 12-13 MAE).
• 21-day forecasts underperform the simple persistence baseline.

Experimental setup

Datasets: Santé Publique France COVID-19 department-level data NOAA department-level weather data Bordeaux University Hospital EHR data (Dryad)

Benchmarks: Bordeaux University Hospital SARS-CoV-2 hospitalization forecast (14-day horizon)

Metrics: MAE MRE MAEB MREB

Baselines: Elastic-net XGBoost LSTM Transformers Informer PatchTST Prophet RC-RS RC-PCA-GA

Builds on: echo state network elastic-net regression PatchTST Informer

Closest papers by meaning

• Dynamics-Based Feature Augmentation of Graph Neural Networks for Variant Emergence Prediction (AAAI 2025 · similarity 0.80)
• A Generalizable Physics-Enhanced State Space Model for Long-Term Dynamics Forecasting in Complex Environments (ICML 2025 · similarity 0.80)
• CGS-Mask: Making Time Series Predictions Intuitive for All (AAAI 2024 · similarity 0.79)
• Spectral-Aware Reservoir Computing for Fast and Accurate Time Series Classification (ICML 2025 · similarity 0.79)
• EARTH: Epidemiology-Aware Neural ODE with Continuous Disease Transmission Graph (ICML 2025 · similarity 0.79)
• Cumulative Difference Learning VAE for Time-Series with Temporally Correlated Inflow-Outflow (AAAI 2024 · similarity 0.79)