How much can models help?

I have data on about 600 asset swaps per year, and track their year ahead change in yield. I then bucketed the data into 20 sections by asset swap level, then sorted within those buckets by the Merton model (ie, a model that treats equity as an option on the value of the firm, the basis of KMV's EDFs). I did this going back to 1999.

You can see than the Merton model correctly predicts the higher risk credit rise in yield by about 150 basis points per year in bear markets (bear market: prices go down, swap spreads go up). But surprisingly, in bull markets, there is less power in the Merton model. This is surprising because default prediction gets worse in a crisis than in good times, that is, standard power metrics are worse trying to explain the 2002 defaults, than the 1998 defaults. But this exercise suggests that in predicting prices, there is more power in a model in the bear markets, than the bull or non-bear markets. I remember the big boom in credit in 2003, and it was insane. Everything rallied, and the crappier names rallied most. You do not want to be short 'crap' when the market is rallying credit, a situation I imagine will happen soon in nonfinancials.

Predicting default is not the same as predicting prices, but of course they are highly correlated objectives. The trick is to find a simple, but strong pattern that others have not.