Exam AWS Certified Machine Learning - Specialty topic 1 question 209 discussion

For time series data, it is important to split the dataset chronologically, with the training dataset containing the earlier dates and the validation dataset containing the later dates

A. YES - it is forecasting, so you want to predict the future and 20% of the data points after a date will do so B. NO - it is forecasting, we want to simulate an actual use case and not predict the past C. NO - there is data leakage as future datapoints are used in the predictions C. NO - there is data leakage as future datapoints are used in the predictions

Time series keep the order

A, As it's a time series problem.

It's a timeseries problem, then the splitting needs to be made by date.

Option A is the recommended approach where the training dataset contains historical prices that precede a certain date, and the validation dataset contains prices that occur after that date. This ensures that the model is trained on past data and evaluated on future data, which is more representative of real-world performance. Option D is NOT the recommended approach for time series data because it ignores the time aspect of the data. Randomly sampling data points without considering the time sequence can result in data leakage and poor model performance.

a since this is time series problem

a https://towardsdatascience.com/time-series-from-scratch-train-test-splits-and-evaluation-metrics-4fd654de1b37

Because it randomly selects data points for both the training and validation datasets, ensuring that the samples are representative of the entire dataset and reducing the chances of overfitting. By randomly sampling without replacement, the data scientist can avoid any biases in the selection of data points and ensure that the training and validation datasets are independent.

I think it should be A!