Regression task: predicting stock prices in the 5th year based on 4 years of data using boosting. What mistake did we make?

The main mistake in the task of predicting stock prices for the 5th year based on 4 years of data using boosting is a violation of the principle of temporal data splitting and potential data leakage.

If the model is trained on data that includes information from the future (for example, data from the 5th year or features dependent on future values), it will be overfitted and show unrealistically high accuracy.

For time series, it is important to use rolling or sequential splitting, where training is only on historical data, and testing is on subsequent periods. Otherwise, the model will not be able to adequately predict the future.

It is also worth considering that stock prices are subject to noise and external factors, so simple boosting without accounting for temporal structure and additional features can lead to poor generalization.