Backtesting Macroprudential Stress Tests

In this paper, we consider models of price-mediated contagion in a banking network of common asset holdings. For these models, the literature proposed two alternative classes of liquidation dynamics: threshold dynamics (banks liquidate their investment portfolios only after they have defaulted), and leverage targeting dynamics (banks constantly rebalance their portfolios to maintain a target leverage ratio). We introduce a one-parameter family of non-linear liquidation functions that interpolates between these two extremes. We then test the capability of these models to predict actual bank defaults (and survivals) in the United States for the years 2008-10. We show that the model performance depends on the type of shock being imposed (idiosyncratic versus systematic). We identify the two most relevant asset classes, for which the model has predictive power when these asset classes are exposed to an initial shock. In these cases, the model performs better than alternative benchmarks that do not account for the network of common asset holdings, irrespective of the assumed liquidation dynamics. We also show how the best performing liquidation dynamics depend on the combination of the initial shock level and the market impact parameter, on the cross-sectional variation in the market impact parameter, and on the number of asset liquidation rounds