New tsbootstrap Library Enhances Time Series Uncertainty Quantification

Researchers have introduced tsbootstrap, a novel Python library designed to improve uncertainty quantification and conformal prediction for time series data. Traditional methods often assume data independence, which is frequently violated in real-world applications like finance and sensor readings. This new tool offers a comprehensive suite of block, residual, sieve, and wild resampling techniques, alongside classical bootstrap confidence intervals and adaptive conformal calibrators. The library's design allows users to select methods via a single, typed API, making it accessible for practitioners. Performance studies indicate that tsbootstrap significantly reduces coverage deficits compared to IID bootstrap methods when dealing with dependent data, particularly with sieve resampling showing near-nominal coverage for short-memory linear dependencies. Furthermore, the compiled backend and streaming reduce capabilities ensure efficient execution, outperforming existing alternatives in speed and memory usage.