This paper presents an investigation into the performance of support vector machine (SVM) in short-term travel-time prediction in comparison with baseline methods, including the historical mean, current time based, and time varying coefficient predictors. To demonstrate the SVM performance, 1-month time-series speed data on a section of Pan-Island Expressway in Singapore were used to estimate the travel time for training and testing the SVM model. The results show that the SVM method significantly outperforms the baseline methods in both normal and recurring congestion over a wide range of prediction intervals. In studying SVM prediction behavior under incident situations, the results show that all the predictors are not responsive enough using 15-minute aggregated field data, but the SVM predicted outcome follows the test data profile closely for 2-minute aggregated simulated data. Finally, to improve the prediction performance, an empirical k-nearest neighbor method is introduced to retrieve patterns closest to the test vector for SVM training. The results show that k-Nearest Neighbor is an attractive tool for SVM travel-time prediction. In retrieving the most similar patterns for SVM training, k-nearest neighbor allows dramatic reduction of training size to accelerate the training task while maintaining prediction accuracy.