This paper applies machine learning algorithms to construct non-parametric, nonlinear predictions of mortgage loan default. I compile a large dataset with over 20 million loan observations from Fannie Mae and Freddie Mac, for the period 2001-2016 at the quarterly frequency. Different machine learning algorithms are applied to predict in sample (training sample), and to forecast out-of-sample (testing data). We find that the forecast performance of nonlinear and non-parametric algorithms are substantially better than the traditional logit model. Additionally, machine learning algorithms allow identification of the predictive power of specific variables. The results indicate that loan age is the most important predictor of loan default before and after the 2008 financial crisis. However, we find that market loan-to-value is the most effective predictor of mortgage loan default during the recent financial crisis. Finally, we use machine learning to formulate risk-based capital stress tests for Fannie Mae and Freddie Mac under different scenarios. We forecast their mortgage credit losses and associated capital needs during the financial crises. The results obtained are more accurate than those from the Federal Housing Enterprise Oversights (OFHEO), and other existing stress test studies.
With the severity of the 2008 financial crisis, and apparent inefficacy of traditional monetary and fiscal policies, the Federal Reserve together with the U.S. government introduced unconventional policy measures. The Large Scale Asset Purchase (LSAP) and Troubled Asset Relief Program (TARP) are some of these policies introduced by the Federal Reserve and Department of Treasury. While these policies may have been important in preventing a deepening of the financial crisis and laying the foundation for the economic recovery, there were collateral effects on bank profitability. In this paper, we study the impact of both the LSAP and TARP programs on banks’ profit and risk taking using a large panel. I consider both LSAP and TARP transactions in the universe of 800 bank holding companies. That is, differently from previous studies, I consider not only long and short interest rates and macroeconomic control variables, but also time varying transactions for each bank. The results indicate that these programs had a positive effect on banks’ profit.
There is a large literature on forecast combination for key economic variables. The idea is to investigate whether combining several models improve the forecasting performance. However, models’ forecast performance might change over time or during different phases of the business cycle. This paper considers this possibility and proposes a Dynamic Conditional Correlation model (DCC) to obtain optimal time-varying combination weights. Several models are used to forecast the U.S. inflation, which are combined using time varying weights estimated by the DCC model. We find that the mean square forecast errors (MSFE) obtained from the forecast combination with time varying weights are substantially lower than the MSFE for each model and for constant weights.
This paper uses a small-scale DSGE model for the economy of Iran to analyze monetary policy. The model is extended to include housing and oil sectors. The model is adapted for the peculiarities of Iran’s Central Bank, which uses money supply as a function of oil income and production growth. The reason is that Iran’s economy does not have market interest rates (interest rates are fixed and determined annually by the Executive Government). We study the reaction function of the model to technology, oil, and monetary shocks in this specific Iranian monetary policy framework. The results show that monetary shocks has only nominal effect on inflation but not on the real sector such as investment, consumption, or production. Also, positive oil income shocks lead to an increase in inflation instead of an increase in production. That is, the Dutch disease mechanism is found for Iran economy, according to the model. This paper also considers and estimates an optimal monetary policy rule. The results are compared using simulation methods.
University of California, Riverside, Economics Department 900 University Ave, Riverside, CA, 92507 firstname.lastname@example.org