Sponsor- and trigger-specific determinants of CAT bond premia: a summary

A CAT bond’s loss distribution is estimated by risk modeling companies such as Applied Insurance Research Worldwide, Risk Management Solutions or EQECAT.

Alternative parameters to describe the loss distribution are the probability of first loss (PFL), conditional expected loss (CEL) and probability of last loss (PLL) (the probability of losses suffered by the CAT bond equal to or greater than the limit) (Lane 2000; Wang 2000, 2004; Dieckmann 2008; Lei et al. 2008; Ahrens et al. 2009).

See, for example, Lane and Beckwith (2017). In the following, we consider the sponsor’s participation in the structuring and placement process and the sponsor’s affiliation with the structuring and placement agent as synonyms.

An underwriter in an ABS deal is comparable to a structuring and placement agent in a CAT bond deal. In the CAT bond literature, these terms are sometimes used synonymously, as, for example, in Lane and Beckwith (2017). The difference between the two functions is that an underwriter engages in a public offering and purchases securities before offering and reselling them to investors, whereas a structuring and placement agent engages in private placements, in which securities are directly offered to investors (Temel 2001). CAT bonds are usually offered through private placements (Cummins and Weiss 2009; Edesess 2015), so that we will subsequently use the term structuring and placement agent (or for brevity placement agent).

The spread is calculated by Lane Financial LLC in three steps. First, bid and ask quotes are surveyed from several dealers. Second, the values are averaged per dealer to obtain a single yield spread per dealer. Third, the quarterly average spreads are averaged again across dealers, yielding the dependent variable of the analysis. The dealers providing secondary market spreads are Aon, BNP Paribas, Cochran Caronia, Goldman Sachs, Lehman Brothers, Merrill Lynch, and Swiss Re. The composition of dealers slightly changed over time (Lane and Beckwith 2006, 2009, 2010).

Following the methodology of the CRAs S&P, Fitch, and Moody’s, we define all ratings above “BBB” or the Moody’s equivalent “Baa2” as investment grade (Moody’s 2016; Standard & Poor’s 2016; Fitch 2017).

Before establishing the model in Eq. 1, we conducted a benchmark analysis to identify the share of the variance in CAT bond premia that can be explained by bond-specific, by sponsor-specific, and by time-dependent variables. For reasons of brevity, the results of that analysis are not presented here but are available from the authors on request. The analysis furnished two important implications. First, year-fixed effects provide a good approximation of the variance in CAT bond premia over time. Second, time effects explain only a minor portion of the variance of CAT bond premia compared to time-invariant bond- and sponsor-specific effects, which leads us to apply random effects estimation to reveal these influencing factors, focusing on the characteristics of the CAT bond sponsor.

For the two models presented in the following, a Breusch-Pagan test is applied to verify that a random effects model is appropriate compared to a pooled OLS model Baltagi and Li (1990). As reported in the bottom of the Table 4, both models reject the test’s null hypothesis of zero variance σ_a of the individual intercept at a high significance level (p < 0.1%). Hence, random effects estimation is more appropriate than pooled OLS estimation.

We also analyzed sponsor characteristics with a more granular set of rating classes in our model. The consideration of dummy variables for the rating classes “AAA”, “AA”, “A”, “BBB”, “BB and worse” and “No rating” showed no systematic differences between rating classes within the investment grade segment or the speculative grade segment.

Comparing models (I.1) and (I.2), it can be noted that the exclusion of the Experience variable leads to a higher adjusted R² but also that the R² is increasing. This result stems from the calculation of goodness of fit. The exhibited R² corresponds to Eq. 1, but the parameter estimation with random effects is based on quasi-demeaned data (i. e., the random effects transformation of Eq. 1). Thus, the calculated R²-values do not have all of the ordinary properties of the OLS R², and the inclusion of additional variables does not necessarily increase the R² (StataCorp 2015; Wooldridge 2015).