Question

When making training, validation, and test sets for predicting future stock prices, what strict rule must be followed about the order of the data?

Accepted Answer

The strict rule that must be followed about the order of data when making training, validation, and test sets for predicting future stock prices is that the data must be strictly chronological. This means that all data points within the training set must occur historically before any data points in the validation set, and all data points within the validation set must occur historically before any data points in the test set. For instance, if you have stock price data from 2000 to 2023, your training set might use data from 2000 to 2021, your validation set from 2022, and your test set from 2023. You cannot mix data from 2023 into the training or validation sets, nor can 2022 data appear after 2023 data in these partitions. A training set is the portion of the available historical data used to teach a machine learning model, enabling it to learn patterns and relationships. A validation set is a separate portion of historical data, chronologically following the training set, used to tune the model’s settings, known as hyperparameters, and to evaluate different model configurations during the development phase. This process helps prevent the model from becoming overly specific to the training data, a problem called overfitting, and allows for objective model selection before final evaluation. A test set is the final, completely independent portion of historical data, chronologically following both the training and validation sets, used only once to provide an unbiased estimate of the final model’s performance on truly unseen data. This simulates how the model would perform when making real-world predictions on future, unknown stock prices. The critical reason for this strict chronological ordering is to prevent a severe issue known as data leakage. Data leakage occurs when information from the future inadvertently becomes available to the model during its training or validation phases. In the context of stock price prediction, if a model were exposed to future stock prices, even indirectly, through a shuffled or improperly ordered dataset, it would effectively “see” the answers beforehand. This would lead to artificially inflated performance metrics that would not be achievable in a real-world scenario where future prices are unknown. By maintaining strict chronological separation, the model only learns from information that would genuinely be available at the time of prediction, ensuring that its performance evaluation is realistic and its predictions are reliable for genuinely unseen future data.

Home → All Courses → Engineering and Technology Courses → DeepLearning.AI TensorFlow Developer Certificate → Flashcard

When making training, validation, and test sets for predicting future stock prices, what strict rule must be followed about the order of the data?