Definition
Overfitting occurs when a model adapts to noise or quirks in its development sample as though they were stable structure. The fitted model can look strong in-sample while generalizing poorly to new observations. Flexible models, many candidate specifications and small samples increase the risk.
How it is calculated
Overfitting is sample-specific fit that does not carry reliably to data that were not used to choose the model.
Interpretation
A model can memorize the fingerprints of its training sample without learning a reusable pattern. A clean holdout asks whether the pattern survives contact with unseen data. Even that check is one piece of evidence, not a universal promise about future samples.
Assumptions and conventions
-
There is a distinction between data used to choose the model and data used to evaluate it.
-
The evaluation process prevents information from the evaluation sample leaking into model selection.
No inherent physical unit; performance metrics retain their own units.
Limitations and common misreadings
-
A high in-sample statistic is weak evidence of generalization by itself.
-
Out-of-sample testing can be overfit too if reused repeatedly.
References
- Trevor Hastie, Robert Tibshirani, Jerome Friedman. (2009). The Elements of Statistical Learning. Springer Series in Statistics. Chapter 7, Model Assessment and Selection. doi:10.1007/978-0-387-84858-7. Source

