Small-Scale Machine Learning Over Scarce and Unreliable Data: Sovereign Credit Grades Prediction

Abstract

According to the Bank for International Settlements, investment banks across the globe now hold more than a trillion US dollars' worth of sovereign bonds. Sovereign credit grades’ movements largely drive these positions' risk and volatility. Their prediction is thus crucial to better manage sovereign credit portfolios. Available economic data is unreliable, scarce and restricted by the number of sovereign entities. Established historical models used to set and predict sovereign credit grades centre around restrictive linear modelling. Modern machine learning techniques require numerous parameters and vast datasets to converge over noisy data. Through thorough data processing, motivated by both economic and statistical insight, we put forward classifiers which novelly demonstrate that small neural networks and random forests can calibrate accurately on poor, lopsided macroeconomic datasets. Their accuracy outperforms all known industry and published implementations, both linear and neural network based. Our results are cross-validated on carefully isolated data, both temporally and geographically, replicating a production situation. This architecture is the first capable of predicting sovereign credit grades’ evolution with accuracy high enough to meet the intuitive and easily implemented “constant” classifier benchmark. Unlocking the use of modern statistical methods to small, low quality economic and financial datasets, our predictors underlie the core importance of extensive problem specific data pre-processing in machine learning for macroeconomic classification.