Prediction-powered inference for time series across space
Researchers adapt prediction-powered inference to give valid intervals for temporally dependent spatial time series.
The paper estimates future expected labels, such as crop yield, at many spatial locations when only a short recent period is labeled and covariates such as weather cover a longer span. Standard prediction-powered inference assumes i.i.d. samples, which fails under temporal dependence, and existing HAC methods do not handle imputed labels. The authors provide bias-corrected point estimates and confidence intervals for this setting and report that the method outperforms natural alternatives.
- Targets short labeled series plus longer unlabeled covariates across locations.
- Corrects imputation bias when i.i.d. prediction-powered inference does not apply.
- Accounts for temporal correlation and heteroskedasticity with imputed labels.
- Reports better point estimates and intervals than natural alternatives.
From the original article
1 min read at arxiv.orgThe following motif is common in spatiotemporal settings: we have a sequence of covariate and label pairs observed for a relatively short, recent time period. We have access to unlabeled covariates over a longer time period. Data is observed over many spatial locations.
By Shahzar Rizvi · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.