Robust neural network with applications to credit portfolio data analysis.

Area of Science:

• Statistics
• Machine Learning
• Data Science

Background:

• Nonparametric conditional quantile estimation is crucial for understanding data distributions.
• Existing methods may struggle with outliers and prediction band construction.
• Neural networks offer flexible modeling capabilities.

Purpose of the Study:

• To propose a novel robust neural network (RNN) method for nonparametric conditional quantile estimation.
• To evaluate the smoothing performance of RNN in the presence of outliers.
• To demonstrate the utility of RNN for constructing prediction bands.

Main Methods:

• Combining quantile regression with neural network architecture.
• Developing a Majorization-Minimization (MM) algorithm for optimization.
• Utilizing Monte Carlo simulations to assess RNN performance.

Main Results:

• The proposed RNN method exhibits good smoothing performance, even with outliers.
• RNN effectively facilitates the construction of prediction bands.
• Simulations and real-data applications show RNN's advantages over local linear regression and regression splines.

Conclusions:

• The robust neural network (RNN) is a powerful tool for nonparametric conditional quantile estimation.
• RNN provides a robust and effective approach for handling outliers.
• This method offers an advantageous alternative to traditional nonparametric regression techniques.