Many hikers enjoy hiking in the low mountains near cities in Japan. However, some hikers are careless and do not prepare or plan their hikes well enough due to the low altitude and complexity of the trails, which can lead to lost routes and accidents. In this study, GPS trajectory data of past hiking are converted into series data in grids and used as training data to construct an LSTM behaviour model that predicts the next grid to which the hiker will move. The next grid series is predicted by a beam search method using the grid movement probability estimated by the constructed model. The predicted grid sequences are compared with the actual movement grids to detect wayfinding errors. Furthermore, the computational efficiency is improved by pruning the grids. The method is validated using actual mountaineering data.