Predicting Fiscal Crises: A Machine Learning Approach

Abstract

In this paper I assess the ability of econometric and machine learning techniques to predict fiscal crises out of sample. I show that the econometric approaches used in many policy applications cannot outperform a simple heuristic rule of thumb. Machine learning techniques (elastic net, random forest, gradient boosted trees) deliver significant improvements in accuracy. Performance of machine learning techniques improves further, particularly for developing countries, when I expand the set of potential predictors and make use of algorithmic selection techniques instead of relying on a small set of variables deemed important by the literature. There is considerable agreement across learning algorithms in the set of selected predictors: Results confirm the importance of external sector stock and flow variables found in the literature but also point to demographics and the quality of governance as important predictors of fiscal crises.

Assessing Macro-Fiscal Risk for Latin American and Caribbean Countries

“…Specifically, the RF yields an average accuracy of 80%, while the logit model is 70%-75% accuracy accurate. In this regard, Hellwig (2021) found that RF approaches consistently outperform heuristic benchmarks and other statistical models. Other studies such as IMF (2021) found that the RF model performed better compared to other ML models in emerging markets and low-income countries, based on the sum of errors and the area under the curve (AUC).…”

“…The leading indicator presented in this paper is based on the nonparametric approach proposed by Kaminsky, Lizondo, and Reinhart (1998) and incorporates calculations of the risk threshold level for each variable using the already proposed ALE technique. Contributions such as those made by Jarmulska (2020) or Hellwig (2021) demonstrate how ML models can be very good at forecasting fiscal stress. However, none of them takes these models to the construction of a leading indicator of macro-fiscal stress using risk thresholds.…”

“…Second, it proposes a leading indicator encompassing fiscal, macroeconomic, and institutional factors by employing both methodologies. Third, it adds to the large and growing literature on EWS for emerging economies while comparing LAC with other emerging and advanced economies (Hellwig, 2021;Jarmulska, 2020;Hernández de Cos et al, 2014).…”

Fiscal Crises: The Role of the Public Debt Investor Base and Domestic Financial Markets as Aggravating and Mitigating Factors

“…Badia et al ( 2020) use machine learning models, specifically random forest algorithms, to predict fiscal crises and suggest, in line with Mullainathan and Spiess (2017), that these algorithms give better out-of-sample predictions by potentially incorporating a very large number of predictors without running into overfitting problems. Hence, as a model for making predictions, machine learning tools tend to outperform traditional approaches such as logit or probit (Hellwig (2021)). Machine learning tools do however have important downsides.…”

Predicting financial crises with machine learning methods

“…The essence of these approaches is to select the most applicable hyperparameters for these machine learning methods. To choose the optimal hyperparameters, recent studies have tended to focus on the out‐of‐sample predictive performance and have used k‐fold cross‐validation to determine the hyperparameters according to the out‐of‐sample predictive accuracy (Chetverikov et al, 2016; Hellwig, 2021; Holopainen & Sarlin, 2017; Jarmulska, 2020). Although simple cross‐validation of the training and test groups can alleviate the overfitting problem to a certain extent, this type of hyperparameter selection still inevitably refers to the information of the test group.…”