A great deal of investment research presents discoveries of statistical associations between assumed explanatory variables and such performance variables as price returns or credit spreads. Many of these results suffer from the problem of misspecified models or spurious correlations. The authors present structural equation modeling (SEM) as a better framework for testing the feasibility of assumed cause/effect relationships. SEM models test how close variance–covariance matrixes are to their theoretical structure under various assumptions of causality. Although the methodology cannot establish causal relationships, it can “reject” or “fail to reject” various causal hypotheses.

The authors use the SEM methodology to establish the plausibility of a causal relationship between credit rating agency (CRA) ratings and credit spreads in the US asset-backed securities (ABS) market. Specifically, three causal hypotheses fail to be rejected using the SEM methodology. In all of these models, a lower rating is associated with a higher spread.

In the secondary market, the authors look at returns across three time windows: one day, three days, and five days after the date of a ratings change. Only at the one-day horizon is there evidence to support the hypothesis that CRA rating upgrades or downgrades lead to positive or negative returns. At the three- and five-day horizons, no evidence supports the hypothesis that rating changes affect returns. It appears that investors turn their attention to the broader stock and bond markets at these horizons.

The authors analyze 24,637 tranches of ABS issued in the United States from December 1999 through December 2015. Variables that could potentially cause changes in the spread include the number of tranches in the structured product, the weighted average maturity of the product, and the collateral (e.g., mortgages, auto loans, and student loans), as well as other variables. Marketwide variables include indexes on the stock, government bond, and corporate bond markets. Credit ratings are collected from the four major CRAs (Moody’s Investors Service, Standard & Poor’s, Fitch Ratings, and DBRS).

The authors investigate several plausible causal relationships, and among those models where the causal relationships are not rejected, the Bayesian information criterion is used to select the “best” model.

The popularity and accessibility of machine learning techniques have made it easier for researchers to uncover statistical relationships between explanatory variables (or, optimistically, alpha) and security returns. Structural equation modeling could be used as a complementary technique to assess the likelihood of causal relationships between variables and returns. The methodology is not foolproof (it is vulnerable to omitted variables bias, for example), but it offers a reality check when used alongside other techniques used in, for example, factor discovery/modeling.