Comments on: Risk per business type http://ozrisk.net/2008/07/22/risk-per-business-type/ Risk Management in Australia Sun, 18 Apr 2010 08:52:21 +0000 hourly 1 http://wordpress.com/ By: Martin Davies http://ozrisk.net/2008/07/22/risk-per-business-type/#comment-26462 Martin Davies Fri, 05 Sep 2008 03:02:34 +0000 http://ozrisk.wordpress.com/?p=348#comment-26462 This is the heart of an expert credit risk system, being able to profile a set of customers within a portfolio to understand the contribution of default to the portfolio part 1. Part 2 would be able to understand what factors have driven default in those sub set of customers, separate out the factors and assign specific ratings so that a score can be generated and used at origination. One important point to note is that such dynamics are not static even throughout a month of a year. Then of course we are only talking about Probability of Default here not LGD/LGE, those would require duration analysis to be carried out, a model of term structure in reality that might be achieved by a polynomial regression process. Obviously people have written books on the above paragraph and not talking about data, matrix regression can be achieved using Gibbs Sampling but the data needs to be filtered using a process such as Multivariate Discriminant Analysis. After this, logical ordering can occur and the final decision tree built up. I have a presentation on the above which is quite basic but eludes to how this can be achieved so please do email me or Clive and I can provide further details. Now back to our matrix: The matrix needs to be multi period which gives it several dimensions but before default adjusted returns can be understood the overall duration of the matrix needs to be dimensioned. After this a Markov Chain process is established which can be used to identify propagation values across the transition matrix. Amazing as it may sound I have a basic example in Microsoft Excel for fixed term notes that also shows recovery percentages across the matrix. Again I can email this to you or please contact Clive but it allows the foundation part of the framework to be built up. It is a starting place which you can build on. Have a pleasant day. This is the heart of an expert credit risk system, being able to profile a set of customers within a portfolio to understand the contribution of default to the portfolio part 1. Part 2 would be able to understand what factors have driven default in those sub set of customers, separate out the factors and assign specific ratings so that a score can be generated and used at origination. One important point to note is that such dynamics are not static even throughout a month of a year. Then of course we are only talking about Probability of Default here not LGD/LGE, those would require duration analysis to be carried out, a model of term structure in reality that might be achieved by a polynomial regression process.

Obviously people have written books on the above paragraph and not talking about data, matrix regression can be achieved using Gibbs Sampling but the data needs to be filtered using a process such as Multivariate Discriminant Analysis. After this, logical ordering can occur and the final decision tree built up.

I have a presentation on the above which is quite basic but eludes to how this can be achieved so please do email me or Clive and I can provide further details.

Now back to our matrix:

The matrix needs to be multi period which gives it several dimensions but before default adjusted returns can be understood the overall duration of the matrix needs to be dimensioned. After this a Markov Chain process is established which can be used to identify propagation values across the transition matrix. Amazing as it may sound I have a basic example in Microsoft Excel for fixed term notes that also shows recovery percentages across the matrix. Again I can email this to you or please contact Clive but it allows the foundation part of the framework to be built up. It is a starting place which you can build on.

Have a pleasant day.

]]>