Similarity search in time-series databases finds such data sequences whose changing patterns are similar to that of a query sequence. For efficient processing, it normally employs a multi-dimensional index. In order to alleviate the well-known dimensionality curse, the previous methods for similarity search apply the Discrete Fourier Transform (DFT) to data sequences, and take only the first two or three DFT coefficients as organizing attributes. Other than this ad-hoc approach, there have been no research efforts on devising a systematic guideline for choosing the best organizing attributes. This paper first points out the problems occurring in the previous methods, and proposes a novel solution to construct optimal multi-dimensional indexes. The proposed method analyzes the characteristics of a target time-series database, and identifies the organizing attributes having the best discrimination power. It also determines the optimal number of organizing attributes for efficient similarity search by using a cost model. Through a series of experiments, we show that the proposed method outperforms the previous ones significantly.