This argument is easy to resolve by doing factor analysis of dataset with (say) CDS and without. Though those studies were done in better (more liquid) times, and there may be objections to results in the current marrkets, they would definitely apply to the timeframe of the paper.

I wasn't trying to suggest that spreads are not important in prediction, but that determining empirically if and how important they are is very difficult.

QuoteOriginally posted by: And2This argument is easy to resolve by doing factor analysis of dataset with (say) CDS and without. Though those studies were done in better (more liquid) times, and there may be objections to results in the current marrkets, they would definitely apply to the timeframe of the paper.I don't see how this would help. We know that there is a difference between spreads and Pds, we don't know why.

QuoteOriginally posted by: gardener3QuoteOriginally posted by: And2This argument is easy to resolve by doing factor analysis of dataset with (say) CDS and without. Though those studies were done in better (more liquid) times, and there may be objections to results in the current marrkets, they would definitely apply to the timeframe of the paper.I don't see how this would help. We know that there is a difference between spreads and Pds, we don't know why.I did not mean that CDS's determine PD's, but having significant (or, even if some) explanatory power, better be included in study of PD's. Ah, I see we go in circles...I am not sure about your example with the wine companies. In ideal world, bonds will be trading at discount, but not necessarily CDS - that is in ideal world they would exactly reflect the DP's. This is probably the main argument for inclusion CDS's in the studies in the first place.

QuoteOriginally posted by: gardener3I wasn't trying to suggest that spreads are not important in prediction, but that determining empirically if and how important they are is very difficult.My beach-time reading today was Kealhofer and Kurbat, 'Predictive Merton models', (Risk, feb 2002) in which theyargue that a good way to test if one measure of default adds something to another (or vice-versa) is an 'intra-cohort' analysis, whichthey clearly explain. Then, they use it to argue that Moody's bond ratings add nothing to the KMV Expected DefaultFrequency (EDF), while the EDF adds significant predictive power to the ratings. Since I see Moody's acquired them in the same monthfor $210 million, the argument must have been pretty persuasive -- or at least very annoying! Anyway, somewhat similar to And2's suggestion, an obvious project for somebody -- if it hasn't been done -- is to attempt the same type of analysis except for EDF vs CDS-implied (Q)-prob of default. Update: A little googling to see what has actually been done turns up this March 2010 study by the (now Moody's) KMV.Abstract:QuoteIn this paper, we present a framework that links two commonly used risk metrics: default probabilities and credit spreads. This framework provides credit default swap-implied (CDS-implied) EDF (Expected Default Frequency) credit measures that can be compared directly with equity-based EDF credit measures. The model also provides equity-based Fair-value CDS spreads (FVS) that can be compareddirectly with observed CDS spreads.CDS-implied EDF credit measures and fair-value spreads are powerful tools that risk managers can use to extend coverage of credit risk measures, enhance the assessment of default risk, and assess the relative value of various credits. With CDS-implied EDF credit measures, we can provide default risk measures for the population of entities without traded equity, such as private firms, subsidiaries ofpublic firms, and sovereigns, based on their CDS. For firms with both EDF credit measures and CDS-implied EDF credit measures, risk managers can use both metrics to enhance their assessments of credit risk at the entity level. That is, by comparing information from both markets in a common metric and understanding the differences, risk managers can gain valuable insights into the credit risk of these entities. By using both measures, they can minimize the model risk of relying on one measure alone and increase predictive power of credit risk measures. Additionally, fair-value spreads can be used for mark-to-market valuation and as well as for portfolio management.

On the moody's study: I am surprised they were able to find so many defaults (but still a tenth of that of firms with EDFs even ignoring the time frame). Looking at the chart, the improvement in prediction seems insignificant. Btw, the difference in P & Q can also be due to sentiment or over/under reaction to news, and even if it is risk, 'risk premium' is an elusive concept. One investing strategy would be to look for "mispricing" in CDSs based on EDFs, which I think is what Kealhoffer is doing these days.