Introduction We compare the CNN model against a simple logistic regression model to learn the benefits of simple models. For extracting features, we use the tsfresh package. We find that logistic regression detects five events compared to six by CNN on an untrained data. It also takes lesser time to train compared to CNN. Method 1.Data selection and pre-processing We have used seismic records from the G- network of the Groningen area for detecting low magnitude earthquakes. It comprises 70 borehole stations, each containing 5 sensors; one is a surface accelerometer and the other four are velocity sensors installed at 50m depth intervals. We use data from 47 events recognized in the KNMI catalog between October 1, 2017 and February 28, 2018. For training and validating, we use the lowest depth velocity sensor at five stations (G19, G23, G24, G29, G67) over the 5-months. For testing, we use a separate four-hour dataset. 2.Feature extraction and analysis Tsfresh lets us identify and extract relevant features using various statistical computations like approximate entropy, skewness, variance, standard deviation etc. It calculated 293 relevant features. We use univariate selection and correlation analysis to find the best possible combination and selected top 20 features. 3.Different models used We have compared the performance of logistics regression and CNN models. The CNN model consists of 3 convolutional layers, each followed by a max-pooling layer. Then we flatten the output of last pooling-layer and pass it into two fully connected layers followed by an output layer which determines whether the input is signal or noise. For CNN, the Adam optimizer is used with binary cross entropy as the loss function. Result Two events were already picked in the KNMI catalogue ( at 00:12:28 and at 00:57:46.) of M1.9 and M2.2 for the test data. Table 1 shows the list of operations from which we calculate the top 20 features. Table 2 shows the number of correct predictions, training time, and the time-stamps of uncatalogued events detected for both models. Both models were able to detect the events listed in the catalogue in addition to other uncatalogued events. Conclusion Simple models are easier to understand, debug, train and interpret than the complex black box models. A detailed study on diverse data is needed to improve our understanding.