We propose reservoir computing (RC), a framework for constructing recurrent neural networks (RNNs) with simple training rule, using ferroelectric tunnel junction (FTJ) crossbar array. Not only the weights of connection between reservoir layer and output layer need to be trained, but also random fixed weights within reservoir layer can be realized by the conductance of FTJs. Random weights within reservoir layers have optimal variability, which is dependent not on the type of the distribution but on the reservoir size. Optimal variability can be attained by device-to-device variability in FTJs together with analog converters on the wordlines and the bitlines. Analog resistive switching characteristics in FTJs can also realize the weights connected to output layer, which need to be trained. Furthermore, device-to-device variability in HfFormula OmittedZrxO2 (HZO) FTJ is extensively investigated. Taking into account the low current operation in FTJs, RC with FTJ crossbar array is a potential candidate for neuromorphic computing system with low power consumption.