In atmospheric time series analysis, where only one record is typically available, subsampling (which works under the weakest assumptions among resampling methods), is especially useful. In particular, it yields large-sample confidence intervals of <i>asymptotically</i> correct coverage probability. Atmospheric records, however, are often not long enough, causing a substandard coverage of subsampling confidence intervals. In the paper, the subsampling methodology is extended to become more applicable in such practically important cases.