The future is in the past? A flexible resampling approach to generate multivariate time series

Synthetic time series generation is an essential tool for robustly exploring different climate scenarios and their impacts. While sophisticated generation methods have been developed in the past, they often rely on physical and statistical assumptions and require extensive data for calibration and parameter estimation. We propose a straightforward method for time series generation based on constrained sampling of observations. This approach preserves the physical consistency between variables and maintains the short temporal structure present in the observation. We apply this procedure to generate temperature, precipitation, incoming solar radiation, and wind speed time series sampled from meteorological station observations. We obtain different sets of synthetic time series by constraining the mean temperature according to future scenarios provided by climate model projections. We show that the sampled time series preserve the multivariate dependence structure observed in both historical data and climate projections. While, by design, the method does not generate daily values beyond the observed range, it can simulate multi-day extremes that exceed those in the observational record, such as longer heatwaves. The approach is flexible and can be applied to other variables with other constraints, provided that a sufficiently long observational time series is available and the constraints are compatible with the observed data. The generation procedure may thus prove useful for studying potential future extremes and help in general downscaling tasks.