High-dimensional time series data need dimension-reduction strategies to improve the efficiency of computation and indexing. In this paper, we present a dimension-reduction framework for time series. Generally, recent data are much more interesting and significant for predicting future data than old ones. Our basic idea is to reduce to data dimensionality by keeping more detail on recent-pattern data and less detail on older data. We distinguish our work from other recent-biased dimensionreduction techniques by emphasizing on recent-pattern data and not just recent data. We experimentally evaluate our approach with synthetic data as well as real data. Experimental results show that our approach is accurate and effective as it outperforms other well-known techniques.