A new index of financial conditions

This 2014 European Economic Review paper by Gary Koop and Dimitris Korobilis develops a time-varying-parameter factor-augmented VAR (TVP-FAVAR) with dynamic model averaging (DMA) to construct a financial conditions index (FCI) that lets both the weights on financial variables and the set of financial variables included evolve over time, while estimating the FCI’s relationship with the macroeconomy jointly rather than in a separate step. The model has two blocks: a financial block in which 18 financial variables (asset prices, volatilities, credit, and liquidity measures, with the S&P500 always included and the other 17 subject to variable selection) load on a latent factor (the FCI) and on contemporaneous macro variables (GDP deflator inflation, unemployment, and real GDP growth) with time-varying loadings, and a macro-FCI VAR block in which the macro variables and the FCI jointly evolve with time-varying VAR coefficients; both sets of parameters follow random walks, and the macro variables enter the financial block only to purge the FCI of current macroeconomic effects. Estimation uses a simulation-free two-step dual Kalman filter (no MCMC), with error covariances updated via exponentially weighted moving averages (decay factors 0.96 for observation equations, forgetting factors 0.99 for state equations), and Dynamic Model Averaging/Selection (DMA/DMS) is applied across the 2^17 = 131,072 possible subsets of the 17 optional financial variables using a Raftery et al. (2010) forgetting-factor approach with alpha = 0.99 (alpha = 1 nesting static Bayesian model averaging). Using quarterly U.S. data from 1970Q1-2013Q3, the resulting FCI begins declining before the onset of the 2007-2009 recession and bottoms out in early 2009, tracks the Chicago Fed National Financial Conditions Index most closely among existing indices while dropping earlier and further in 2008, and is built from a DMA-selected subset that averages between about 5 and 8 of the 17 optional variables at any point in time, with substantial switching in which variables are included as conditions change. In out-of-sample forecasting evaluated over 1990Q1-2013Q3 (Table 2) and against existing FCIs over 2000Q1-2013Q3 (Table 3), the TVP-FAVAR-DMA specification with (kappa=0.96, alpha=0.99) delivers lower mean squared forecast error and higher average predictive likelihood than a VAR benchmark at every horizon h=0-4 for inflation, unemployment, and output (e.g., relative MSFE of 0.63 and 0.55 for unemployment at h=0 and h=2), and “TVP-FAVAR-DMA and TVP-FAVAR-DMS almost always forecast better than the TVP-VAR-FCI4” (the TVP-VAR augmented with the Chicago Fed’s index, the best-performing existing-FCI comparator); the authors attribute the largest gains to allowing for stochastic volatility in the model’s parameters, with time-varying VAR coefficients, the FAVAR factor structure, and DMA/DMS each contributing further, smaller improvements.

Gary Koop | Macro Paper Warehouse

A new index of financial conditions