Commonly used thin reservoir prediction methods are highly dependent on drilling data. They have high predictive precision for development areas which have more drilling wells and more homogeneous distribution. Meanwhile, it is difficult to meet the requirement of lithological reservoir prediction in exploration areas which have lower degree of exploration, less drilling wells and inhomogeneous distribution. Predicting the 10–15 m thick reservoir accurately in deep layers is one of the technical bottleneck of lithological reservoirs prediction in Triassic and Permian of Junggar basin. A thin reservoir prediction method based on broadband seismic information is proposed in this paper: firstly, through the forward analysis of logging and seismic spectrum, the frequency of the original seismic data is extended reasonably by continuous wavelet transform. Secondly, the damping factor is used to remodel the spectrum characteristics of extension frequency data, the high and low frequency data which have the same spectrum characteristics to the seismic data and logging are selected. Finally, frequency division iterative inversion is carried out to the selected high and low frequency data to meet the requirement of thin reservoir prediction in the areas which have low degree of exploration and less logging data. This method can use the vertical and lateral information of seismic data effectively. Meanwhile, the high and low frequency iterative inversion can improve the problem of amplitude preservation of seismic data become worse during extension frequency. This method predicts the over 10 m thick reservoir in Permian accurately in Bei43 well area in eastern junggar basin, the predicted agreement rate is more than 83% and provide valuable support for well proposal.