Search Results

You are looking at 1 - 10 of 12 items for :

Clear All
International Monetary Fund. Strategy, Policy, & Review Department

-learning models regularly outperform classical econometric methods . Tree-based models are the most successful in out-of-sample prediction for the financial and fiscal sectors. For the external sector, the signal extraction approach is most effective for sudden stops and Exchange Market Pressure (EMP) events in advanced economies (AEs), while a RF model is better for EMP events in emerging markets (EMs) and low-income countries (LICs). Pooling all countries improves the performance of fiscal and financial models . Pooling all country groups typically improves forecasting for

Klaus-Peter Hellwig
Can countries improve their business climate through reforms in specific policy areas? Kraay and Tawara (2013) find that the answer depends on how we measure the business climate. When regressing seven different business climate indices on 38 policy indicators, they find little agreement among the seven models as to which of those policy indicators matter most. I revisit this puzzle using the same data but replacing their linear models with a Random Forest algorithm. I find a strong consensus across models on the importance ranking of policy indicators: No matter which business climate index is considered, the top ten contributors to a better business climate always include high recovery rates in insolvency proceedings (i.e., cents on the dollar for creditors), shorter border formalities for both importers and exporters, and low costs for starting a business. I show that the marginal effect of reforms is heterogeneous across countries and document how reform priorities depend on country specific circumstances.
International Monetary Fund. Strategy, Policy, & Review Department
The IMF’s Vulnerability Exercise (VE) is a cross-country exercise that identifies country-specific near-term macroeconomic risks. As a key element of the Fund’s broader risk architecture, the VE is a bottom-up, multi-sectoral approach to risk assessments for all IMF member countries. The VE modeling toolkit is regularly updated in response to global economic developments and the latest modeling innovations. The new generation of VE models presented here leverages machine-learning algorithms. The models can better capture interactions between different parts of the economy and non-linear relationships that are not well measured in ”normal times.” The performance of machine-learning-based models is evaluated against more conventional models in a horse-race format. The paper also presents direct, transparent methods for communicating model results.
Jelle Barkema, Mr. Mico Mrkaic, and Yuanchen Yang
This paper dives into the Fund’s historical coverage of cross-border spillovers in its surveillance. We use a state-of-the-art deep learning model to analyze the discussion of spillovers in all IMF Article IV staff reports between 2010 and 2019. We find that overall, while the discussion of spillovers decreased over time, it was pronounced in the staff reports of some systemically important economies and during periods of global spillover events. Spillover discussions were more prominent in staff reports covering advanced and emerging market economies, possibly reflecting their role as sources of global spillovers. The coverage of spillovers was higher in the context of the real, financial, and external sectors. Also, countries with larger economies, higher trade and capital account openess and lower inflation are more likely to discuss spillovers in their Article IV staff reports.
Ms. Marialuz Moreno Badia, Mr. Paulo A Medas, Pranav Gupta, and Yuan Xiang
With public debt soaring across the world, a growing concern is whether current debt levels are a harbinger of fiscal crises, thereby restricting the policy space in a downturn. The empirical evidence to date is however inconclusive, and the true cost of debt may be overstated if interest rates remain low. To shed light into this debate, this paper re-examines the importance of public debt as a leading indicator of fiscal crises using machine learning techniques to account for complex interactions previously ignored in the literature. We find that public debt is the most important predictor of crises, showing strong non-linearities. Moreover, beyond certain debt levels, the likelihood of crises increases sharply regardless of the interest-growth differential. Our analysis also reveals that the interactions of public debt with inflation and external imbalances can be as important as debt levels. These results, while not necessarily implying causality, show governments should be wary of high public debt even when borrowing costs seem low.
Katharina Bergant, Miss Anke Weber, and Andrea Medici
Using micro-data from household expenditure surveys, we document the evolution of consumption poverty in the United States over the last four decades. Employing a price index that appears appropriate for low income households, we show that poverty has not declined materially since the 1980s and even increased for the young. We then analyze which social and economic factors help explain the extent of poverty in the U.S. using probit, tobit, and machine learning techniques. Our results are threefold. First, we identify the poor as more likely to be minorities, without a college education, never married, and living in the Midwest. Second, the importance of some factors, such as race and ethnicity, for determining poverty has declined over the last decades but they remain significant. Third, we find that social and economic factors can only partially capture the likelihood of being poor, pointing to the possibility that random factors (“bad luck”) could play a significant role.
International Monetary Fund. Strategy, Policy, & and Review Department
A careful review has revealed significant scope to modernize and better align the MAC DSA with its objectives and the IMF’s lending framework. This note proposes replacing the current framework with a new methodology based on risk assessments at three different horizons. Extensive testing has shown that the proposed framework has much better predictive accuracy than the current one. In addition to predicting sovereign stress, the framework can be used to derive statements about debt stabilization under current policies and about debt sustainability.
Klaus-Peter Hellwig

) from the original sample. The heterogeneity in variables is achieved by limiting the set of variables available to the algorithm to a small number m try of variables drawn randomly from the full set of variables. This random subset is redrawn at each split. On the one hand, this randomization implies that individual trees are poor representations of the data generating process. However, the diversity among the trees implies that model errors tend to offset each other. Hence, if the number of trees is sufficiently large, RF models can achieve predictive performance

Jelle Barkema, Mr. Mico Mrkaic, and Yuanchen Yang

and easiest models for textual analysis. Therefore, we use LR as our baseline model, complemented by the support-vector-machine (SVM) model and the random forest (RF) model. The LR is a probabilistic classifier that relies on supervised machine learning. Its goal is to train a classifier that can make a binary decision about the class of a new input observation, which in our case is to decide whether a paragraph is about spillovers or not. Consider an input paragraph x, which is typically vectorized and represented as [x 1 , x 2 ,..., x n ]. The classifier output

Ms. Marialuz Moreno Badia, Mr. Paulo A Medas, Pranav Gupta, and Yuan Xiang

-of sample predictive performance of four RF estimated using the features selected by each algorithm against the RF estimated with the full set of variables. We check the statistical significance of the difference between the performance of each algorithm and the full RF model by calculating t-tests based on standard errors adjusted for two-way clustering (see Cameron, Gelbach, and Miller 2011 ). The main performance measure for these comparisons is the area under the receiver-operator curve (AUROC), although other measures such as of the log likelihood and mean squared