Since the limitation of inefficient use and unintelligent frequency selection of the HF (High Frequency) band, a method of HF cognitive frequency selection using Hidden Markov Model (HMM) was proposed. Applying cognitive radio principles to HF communications, HF legacy users were considered as primary users, and the HF radio using cognitive technologies were seen as the secondary user. Firstly, the HMM was established to predict channel states of HF legacy users based on the history data of spectrum sensing; secondly, channel parameters were estimated if the predicted state was idle; finally, the optimal frequency was selected among the channels whose predicted states were idle according to the estimated channel parameters. Simulation results show that the proposed method can be used to actually predict HF legacy users' channel states and quickly estimate channel parameters. Under the given simulation conditions, the successful transmission ratio of the proposed method is 5.54% and 10.56% higher than the methods of random channel selection using HMM prediction and energy detection, therefore the proposed method can select the optimal channel.