New Evaluation Metric CSE Proposed for Irregular Time Series Forecasting
Researchers have introduced Continuous-time Squared Error (CSE), a new evaluation metric designed to address biases in irregular time-series forecasting. Standard benchmarks rely on mean squared error (MSE), but its results are influenced by timestamp sampling distributions, not just model predictions. CSE uses importance weighting to neutralize this sampling effect, offering a fairer assessment of continuous-time predictive performance. The team mathematically proved CSE's asymptotic estimation error is no larger than MSE's. They also developed a benchmark using synthetic, semi-synthetic, and eight real-world datasets to validate the metric. The work is detailed in a paper on arXiv with identifier 2608.17293.
Key facts
- The paper is arXiv:2608.17293.
- It focuses on evaluation metrics for irregular time-series forecasting.
- Existing benchmarks typically use MSE.
- MSE can be biased due to sample-specific timestamp sampling distributions.
- The paper proposes CSE (Continuous-time Squared Error).
- CSE uses importance weighting to remove timestamp sampling influence.
- Theoretical proof: CSE's asymptotic estimation error is no greater than MSE's.
- A benchmark was constructed covering synthetic, semi-synthetic, and eight real-world datasets.
Entities
—