A comparison of ML Models for Bankruptcy Prediction based on performance and interpretability (with Dr. Amit Das)

Abstract: Even when machine learning systems display a high level of performance, the inability of domain experts to understand how they work – how they arrive at their answers – can inhibit the adoption of such systems in organizations. This opacity leads to a lingering suspicion that the performance on one set of data set might not be repeated on the next set of data. Opacity also makes it difficult to justify the system to various stakeholders and raises accountability issues when it fails, i.e. it makes incorrect decisions.

We apply three machine learning algorithms – logistic regression, decision trees, and neural networks –to the problem of predicting corporate bankruptcy with a large set of independent variables. All three algorithms perform comparably well in classification performance (F1-score, Matthews correlation coefficient, and area under the ROC), though they weigh the independent variables differently to arrive at their conclusions.

The three models, each with performance data and the set of top predictors (using Shapley values), are presented to domain experts: university finance professors. We ask them to assess each model regarding the theoretical justification for its selected predictors. A model that uses more justifiable predictors is more interpretable than others. We compare the interpretability ordering of the algorithms to their performance order. Finally, we run the three algorithms using only variables shortlisted by human experts; these models represent a trade-off between performance and interpretability.

Keywords: Human Understanding, Machine Learning, Bankruptcy

Exchange Rate Uncertainty and South Africa’s Bilateral Trade: A Gravity Model Approach (with Dr. Devleena Majumdar)

Abstract: The paper attempts to analyse the impact of exchange rate uncertainty on South Africa’s gross bilateral trade flows between its BRIC partners along with random set of four countries (chosen from G20 countries namely Indonesia, Japan, Mexico, USA) through a gravity model. Categorizing the industries into six segments namely Agricultural Raw Materials, Fuels, Ores and Metals, Chemicals, Machinery and Transport Equipment and All other commodities, over a time span of 19 years (2000-2019), a panel data technique has been applied to examine the effect. The empirical result suggests that even though uncertainty reduces the bilateral trade flows between the four random trading partners, however does not affect the trade flows between the BRIC partners. Further, following the gravity model of trade, the study suggests that the volume of bilateral trade between South Africa and its BRIC and random partners are proportional to the product of each country’s economic growth and distance with an exception of agricultural raw materials for the random trading partners.

Keywords: Gravity Model, Exchange Rate Uncertainty, BRICS, South Africa