2024-06-10-美联储-因素选择与结构断裂（英）_47页_716kb

Summary

Factor selection in asset pricing models is a critical task, complicated by time-varying shifts in relevant risk factors. This paper presents a Bayesian method to detect multiple structural breaks in factor sets using an exhaustive grid search. Using Fama-French six-factor data from 1963 to 2023, the methodology identifies three breaks in 1975 (oil shocks), 1995 (tech boom), and 2005 (pre-GFC). Prior to 2005, dense factor models with five or six factors performed best, but since then, only two (market, profitability) are selected. Ignoring breaks leads to overestimation of dense models. Factor risk premia and market prices of risk vary by regime. The findings highlight the need to account for structural breaks in factor selection to avoid bias and improve investment strategies.

Key Findings

1. Factor Selection Dynamics:

   • Before 2005: Five or six factors selected, with dense models preferred.
   • Post-2005: Shift to parsimony; only two factors (MKT, profitability) selected.

2. Structural Breaks:

   • Breaks at 1975 (inflation, oil shocks), 1995 (tech revolution), 2005 (pre-GFC).
   • Breaks identified by higher marginal likelihoods in Bayesian model scanning.

3. Risk Premia:

   • Time-variation in all six factor risk premia.
   • Equity premium highest in recent regimes (≈10%), value premium declined since disselection.

4. Pricing Performance:

   • Selected factors price all omitted factors in each regime.
   • Tangency portfolio using MKT and profitability yields a Sharpe ratio of 2.74, higher than traditional models (e.g., CAPM, FF3, FF5).

5. Empirical Implications:

   • Factor zoo literature often misattributes factors due to ignoring structural breaks.
   • Post-2005 reforms emphasize parsimony, correcting for over-parameterization bias.