Research

Working papers

Overview: Does fiscal austerity influence sovereign debt spreads in a state-dependent way? To provide a structural investigation, we build an endogenous sovereign default model with long-maturity debt, fiscal rule, and investment. Our model matches Greek default data and provides novel findings. First, it rationalizes the empirical evidence that the response of sovereign debt spreads is state-dependent: under high financial stress, fiscal austerity leads to higher spreads because adverse effects such as economic slowdown outweigh the benefits from reduced debt accumulation. Under low financial stress, fiscal austerity can decrease sovereign spreads and hence alleviates debt crisis. Second, the interplay of default risk and debt accumulation plays an essential role in determining the state-dependence, which is also amplified by capital movement. Third, the likelihood of self-defeating fiscal austerity is nontrivial, especially in the short-run and during economic downturns. Therefore, when conducting a preliminary assessment of the market impact of fiscal policy in small open economies, it is important to consider this state-dependence.


Overview: This paper documents the evidence of nonlinear negative relationships between potential GDP growth forecasts and government debt spreads during the Eurozone sovereign debt crisis (2009-2016). Existing equilibrium default models assuming full information rational expectation (FIRE) on trend growth struggle to explain the new evidence because trend growth is not significantly correlated with spreads. In this paper, we build a new sovereign default model where knowledge of trend growth is imperfect and hence agents learn about it to make optimal decisions. The results of our model show that embedding learning mechanism provides an easy solution to rationalize the above-mentioned new evidence in an equilibrium default model.

 

Overview: Solving dynamic equilibrium economic models can be time-consuming. Researchers have been exploring efficiency improvements through parallelization techniques. In this study, we leverage Matlab Executable (Mex) to implement compiled and parallelized codes executed on both CPU and GPU. The paper provides three key takeaways: First, this platform accelerates model-solving by up to 40 times compared to their serial scripted Matlab code counterparts, surpassing the performance of naive parallel methods such as ’parfor’ and ’gpuArray’ by up to 25 times. Second, the compiled codes are automatically generated from Matlab scripts, eliminating the need for users to have prior knowledge of C or CUDA programming. Third, we observe that the relative performance on CPU and GPU depends on the algorithm and workload size, influenced by architectural differences. In summary, while parallelization based on Mex remains under-discussed in economic literature, it offers a compelling fusion of the convenience of high-level Matlab programming with the efficiency of low-level C and CUDA.

Publications