In this paper, we propose an optimized delay diversity (ODD) scheme for fading intersymbol interference (ISI) channels. The novel scheme requires knowledge of the channel impulse response (CIR) autocorrelation matrix at the transmitter, but the CIRs themselves have to be available only at the receiver side. We derive a cost function for optimization of the ODD transmit filters and we provide a steepest descent algorithm for iterative calculation of the filter coefficients. In addition, an upper bound on the cost function is derived and employed to prove the asymptotic optimality of the generalized DD (GDD) scheme in [1] for very high signal-to-noise (SNR) ratios and transmit filters of maximum length. However, for SNRs of practical interest and reasonable filter lengths the novel ODD scheme significantly outperforms GDD for both optimum and suboptimum equalization.