Structural breaks, model uncertainty and factor selection

This paper addresses the non-standard problem of detecting multiple structural breaks when the data-generating model in each regime is uncertain, with an application to factor selection in empirical asset pricing. Detection is based on the marginal likelihood of break points, obtained by a novel integration over all possible pairings of models across regimes, from all possible models within regimes. The optimal break points maximize this marginal likelihood. Applying this method to the six Fama–French factors on monthly data from 1963–2023, the analysis identifies three breaks – 1982, 1998, and 2009 – and a shift toward more parsimonious models after 1998. Before 1998, five or six factors are selected, but two afterward. Thus, with breaks, there is a move to parsimony, which has implications for the factor zoo literature. Moreover, within each regime, all omitted factors are spanned by the ones selected. Incorporating breaks also leads to substantially different weight allocations in the maximum Sharpe ratio risk factor portfolio.