Residues in protein sequences can be classified into two (exposed / buried) or three (exposed/intermediate/buried) states according to their relative solvent accessibility. Markov chain model (MCM) had been adopted for statistical modeling and prediction. Different orders of MCM and classification thresholds were explored to find the best parameters. Prediction results for two different data sets and different cut-off thresholds were evaluated and compared with some existing methods, such as neural network, information theory and support vector machine. The best prediction accuracies achieved by the MCM method were 78.9% for the two-state prediction problem and 67.7% for the three-state prediction problem, respectively. A comprehensive comparison for all these results shows that the prediction accuracy and the correlative coefficient of the MCM method are better than or comparable to those obtained by the other prediction methods. At the same time, the advantage of this method is the lower computation complexity and better time-consuming performance.