© 2021, The Author(s), under exclusive licence to Springer Nature Switzerland AG.To minimize catastrophic impacts of earthquakes, many researchers have directed their efforts to developing novel computational techniques and models that forecast natural disasters based upon earthquake catalogues, anomalies in the ionosphere, and radon time series databases, etc. In this article, we propose a methodology that categorizes the soil radon gas (SRG) concentration into seismically active and non-active time series by taking the advantage of a stacking ensemble approach and utilizing an automatic anomaly indication function as a post-processing technique. The main idea behind the proposed methodology is of two layers. The first layer uses a stacking ensemble-based approach that incorporates three learners, i.e. a generalized linear model, linear regression and K-nearest neighbors, to train on seismically active and inactive periods to predict SRG concentration. These predictions are then combined with the labeled anomaly data to train a meta-learner, i.e., support vector machine with a radial kernel, that categorizes the series into active and non-active radon time series data. In the second layer, these classifications are then passed to an automatic anomaly indication function that further labels the time series in a group of readings where the level of received indications is greater than or equal to the indication factor. Radon time series (RTS) data has been recorded over a fault line present in Muzaffarabad, a city in the Pakistan territory of Kashmir, ranging from 1 March 2017 to 28 February 2018. Four seismic events occurred during the study period. The conclusion of the study reveals that the proposed methodology accurately locates the anomaly in the RTS data at different window sizes, i.e. with respect to individual days. The evaluation is based purely on the classification after processing of RTS data by both layers of the proposed methodology.