Dimitri Bianco, VP of Quantitative Risk & Research for Agora Data, recently finished his latest research paper Kurtosis Residual Conjecture: A Metric for Optimal Decision Tree Fitting and Beyond. This paper proposes a better method for determining over and under fitted models. The topic could apply to all mathematical models but the focus in this instance is on decision trees.
This paper delves into how to determine a mathematical model’s performance based on the kurtosis of the residuals. Overfitting is when a model performs exceptionally well on the training data but poorly on new data. The model relies too heavily on the originally inputted data and cannot interpret the pattern well enough make accurate predictions. Underfitting is when a model performs poorly on both training and new data. The model is too simple to capture the data’s underlying pattern.
A balance between the extremes of over and under fitted models creates a model that can generalize unseen data sufficiently. A model with residuals closer to a mesokurtic distribution indicates that most of the true information is being extracted from the available data while avoiding a model that is picking up excessive noise in the data.
Improvements to methods in selecting and building more robust models are critical to the success and advancement of the fintech field. By building more accurate models, our loan originators can make safer and sounder decisions. That is one of the reasons Agora has invested in building out a diverse quant team.
Check out SSRN (Social Science Research Network) for a free copy of the full, enlightening read by Dimitri Bianco.
About Agora Data, Inc.
Agora Data, Inc. is a leading consumer fintech company revolutionizing financing for the subprime and non-prime market. Loan originators can secure capital, obtain actionable loan performance data to improve their portfolios, and use a wide range of solutions to grow their business safely. Powered by patent pending artificial intelligence (AI) and machine learning (ML) technology, loan originators can access real-time data analytics and planning resources to help improve the performance of their portfolios. Agora Data, with deep roots in automotive, made history by closing the first-ever asset pooled non-prime auto securitization in 2020 and continues to bring groundbreaking financing solutions to an underserved market. For more information, visit www.agoradata.com or contact us at 1-877-592-4672.