Frame error rate (FER) prediction in wireless communication systems is an important tool with applications to system level simulations and link adaptation, among others. Although in realistic communication scenarios it is expected to have codewords of different lengths, previous work on FER prediction marginally treated the dependency of the FER on the codeword length. In this paper, we present a method to estimate the FER using codewords of different length. We derive a low complexity FER estimator for frames of different length transmitted over a binary symmetric channel of unknown error probability. We extend this technique to coded systems by the use of effective SNR FER predictors. The proposed estimation scheme is shown to outperform other simpler estimation methods.