Cybersecurity has become one of the focuses of organisations. The number of cyberattacks keeps increasing as Internet usage continues to grow. An intrusion detection system (IDS) is an alarm system that helps to detect cyberattacks. As new types of cyberattacks continue to emerge, researchers focus on developing machine learning (ML) based IDS to detect zero-day attacks. Researchers usually remove some or all attack samples from the training dataset and only include them in the testing dataset when evaluating the performance of an IDS on detecting zero-day attacks. Although this method may show the ability of an IDs to detect unknown attacks; however, it does not reflect the long-term performance of the IDS as it only shows the changes in the type of attacks. In this paper, we focus on evaluating the long-term performance of ML based IDS. To achieve this goal, we propose evaluating the ML-based IDS using a dataset that is created later than the training dataset. The proposed method can better assess the long-term performance of an ML-based IDS, as the testing dataset reflects the changes in the type of attack and the changes in network infrastructure over time. We have implemented six of the most popular ML models that are used for IDS, including decision tree (DT), random forest (RF), support vector machine (SVM), na\"ive Bayes (NB), artificial neural network (ANN) and deep neural network (DNN). Our experiments using the CIC-IDS2017 and the CSE-CIC-IDS2018 datasets show that SVM and ANN are most resistant to overfitting. Besides that, our experiment results also show that DT and RF suffer the most from overfitting, although they perform well on the training dataset. On the other hand, our experiments using the LUFlow dataset have shown that all models can perform well when the difference between the training and testing datasets is small.