The fundamental idea of the holdout approach is to avoid overfitting by exposing the model to the new dataset compared to the one that was used for training. Initially, it results in a model accuracy that is below the minimum threshold; however, the method provides an unbiased learning performance and accurate estimate. The dataset is randomly broken down into three subsets in this method:

• Training set: This is a dataset subset used for predictive model building.
• Validation set: This is a dataset subset used to evaluate the performance of the model built in the training phase. It provides a test platform to fine-tune the parameters of a model and select the best performing model. Not every modeling algorithm requires a set of validations.
• Test set: This is a subset of a dataset used to predict the potential future output of a product.