With direct-touch interaction, users are sensitive to very low levels of latency, in the order of a few milliseconds. Assessing the end-to-end latency of a system is thus becoming an important part of touch-devices evaluation, and this must be precise and accurate. However, current latency estimation techniques are either imprecise, or they require complex setups involving external devices such as high-speed cameras.In this paper, we introduce and evaluate a novel method that does not require any external equipment and can be implemented with minimal efforts. The method is based on short-term prediction of the finger movement. The latency estimation is obtained on the basis of user calibration of the prediction to fully compensate the lag. In a user study, we show that the technique is more precise than a similar "low overhead'' approach that was recently presented.