Overfitting Due to Trend Feature

[DryBoss]

The model is overfitting due to the inclusion of the trend feature, which is derived from the entire dataset, including the target variable. This results in misleading performance metrics as the model effectively "memorizes" these features instead of learning generalizable patterns.

The trend feature is calculated using the full dataset, including the target variable (cleaned_cases). The dataset is then split, and the model is trained with these features. Consequently, the model predicts almost perfectly because the trend and seasonal_strength features closely correlate with the target variable.

Note: The trend feature is essentially a processed version of the target variable cleaned_cases.

---

Problem

• The trend feature is derived from the entire dataset, including the target variable (cleaned_cases), causing data leakage.
• This leads to overfitting, where the model learns the target values too easily, resulting in biased performance metrics.

---

Suggested Solution

1. Remove the trend feature entirely, OR
2. Calculate these feature only on the training set to avoid exposing the test set to information about the target variable.
3. Reassess model performance after applying these corrections to obtain a realistic evaluation of its predictive power.

---

Impact

The current preprocessing approach leads to:

• Overfitting, as the model relies heavily on features derived from the target variable.
• Biased performance metrics that misrepresent the model's ability to generalize to new data.

By addressing this issue, the model will provide more reliable and meaningful predictions.

[Ashfinn]

Thanks for the feedback.. You can create a PR mentioning this issue with all the fixes.