Time‐varying fading channels present a major challenge in the design of wireless communication systems. Adaptive schemes are often employed to adapt the transmission parameters to receiver‐based estimates of the quality of the channel. We consider a pilot‐based adaptive modulation scheme without the use of a feedback link. In this scheme, pilot tones (known by sender and receiver) are periodically sent through the channel for the purpose of channel estimation and coherent demodulation of data symbols at the receiver. We optimise the duration and power allocation of these pilot symbols to maximise the information‐theoretic achievable rates using binary signaling. We analyze four transmission policies and numerically show how optimal training in terms of duration and power allocation varies with the channel conditions and from one transmission policy to another. We prove that for a causal estimation scheme with flexible power allocation, placing all the available power on one pilot is optimal. Copyright © 2016 John Wiley & Sons, Ltd.