This study applies state-of-the-art machine learning (ML) techniques to forecast IMF-supported programs, analyzes the ML prediction results relative to traditional econometric approaches, explores non-linear relationships among predictors indicative of IMF-supported programs, and evaluates model robustness with regard to different feature sets and time periods. ML models consistently outperform traditional methods in out-of-sample prediction of new IMF-supported arrangements with key predictors that align well with the literature and show consensus across different algorithms. The analysis underscores the importance of incorporating a variety of external, fiscal, real, and financial features as well as institutional factors like membership in regional financing arrangements. The findings also highlight the varying influence of data processing choices such as feature selection, sampling techniques, and missing data imputation on the performance of different ML models and therefore indicate the usefulness of a flexible, algorithm-tailored approach. Additionally, the results reveal that models that are most effective in near and medium-term predictions may tend to underperform over the long term, thus illustrating the need for regular updates or more stable – albeit potentially near-term suboptimal – models when frequent updates are impractical.

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. Fiscal variables appear to have less predictive value, and public debt matters only to the extent that it is owed to external creditors.

This paper studies the determinants of repeated use of Fund-supported programs in a large sample covering virtually all General Resources Account (GRA) arrangements that were approved between 1952 and 2012. Generally, the revolving nature of the IMF’s resources calls for the temporary sup-port of member countries to address balance of payments problems while repeated use has often been viewed as program failure. First, using probit models we show that a small number of country-specific variables such as growth, the current account balance, the international reserves position, and the institutional framework play a significant role in explaining repeated use. Second, we discuss the role of IMF-specific and program-specific variables and find evidence that a country’s track record with the Fund is a good predictor of repeated use. Finally, we conduct an out-of-sample forecasting exer-cise. While our approach has predictive power for repeated use, exact forecasting remains challenging. From a policy perspective, the results could prove useful to assess the risk IMF programs pose to the revolving nature of the Fund’s financial resources.

Machine learning tools are well known for their success in prediction. But prediction is not causation, and causal discovery is at the core of most questions concerning economic policy. Recently, however, the literature has focused more on issues of causality. This paper gently introduces some leading work in this area, using a concrete example—assessing the impact of a hypothetical banking crisis on a country’s growth. By enabling consideration of a rich set of potential nonlinearities, and by allowing individually-tailored policy assessments, machine learning can provide an invaluable complement to the skill set of economists within the Fund and beyond.

Forecasting macroeconomic variables is key to developing a view on a country's economic outlook. Most traditional forecasting models rely on fitting data to a pre-specified relationship between input and output variables, thereby assuming a specific functional and stochastic process underlying that process. We pursue a new approach to forecasting by employing a number of machine learning algorithms, a method that is data driven, and imposing limited restrictions on the nature of the true relationship between input and output variables. We apply the Elastic Net, SuperLearner, and Recurring Neural Network algorithms on macro data of seven, broadly representative, advanced and emerging economies and find that these algorithms can outperform traditional statistical models, thereby offering a relevant addition to the field of economic forecasting.

Macroeconomic analysis in Lebanon presents a distinct challenge. For example, long delays in the publication of GDP data mean that our analysis often relies on proxy variables, and resembles an extended version of the “nowcasting” challenge familiar to many central banks. Addressing this problem—and mindful of the pitfalls of extracting information from a large number of correlated proxies—we explore some recent techniques from the machine learning literature. We focus on two popular techniques (Elastic Net regression and Random Forests) and provide an estimation procedure that is intuitively familiar and well suited to the challenging features of Lebanon’s data.

We develop a framework to nowcast (and forecast) economic variables with machine learning techniques. We explain how machine learning methods can address common shortcomings of traditional OLS-based models and use several machine learning models to predict real output growth with lower forecast errors than traditional models. By combining multiple machine learning models into ensembles, we lower forecast errors even further. We also identify measures of variable importance to help improve the transparency of machine learning-based forecasts. Applying the framework to Turkey reduces forecast errors by at least 30 percent relative to traditional models. The framework also better predicts economic volatility, suggesting that machine learning techniques could be an important part of the macro forecasting toolkit of many countries.