Here we go:
The industry standard for retail credit models is a scorecard approach. For a very detailed book on how these models are built see "Credit Risk Scorecards" by Naeem Siddiqi (This is 200 or so pages so I'll try to give a summary here).
Simplistically, the way these models work is by having a set of factors (e.g. age, qualification, income, loan to value etc.). Each of these factors is divided into groups called buckets (e.g for age the buckets might be 18-25, 25-35 etc). Then each of these buckets is assigned a score, usually with higher scores indicating lower risk. To get a person's final score, their score for each factor in the model is added together. There is no standard range for these scores but bank's will usually have a cut of below which they will not lend money.
As you can see, using a scorecard model is very easy, hence their popularity for retail credit risk modelling.
Building the scorecard is quite a lot more complicated though. And will follow the process below (more or less):
- Start with a long list of potential factors
- Bucket the factors
- Discard factors that are not appropriate
- Pick the combination of factors that gives the "best" model
- Decide on the location and scale of the scores
- Create the scorecard
- (Optional) Map the scores to a Probability of Default
The long list of factors will generally be either all of the data that is available or it will be a subset of that that is decided on by the modeller and credit risk experts.
Some factors are handily pre-bucketed e.g. qualification. Others are not. The ones that are not are bucketed by maximising a measure called the Information Value. This measures how much information the factor give about payment versus non-payment. This is usually done iteratively but there isn't really a simple way to do it so it often comes down to gut feel and "good enough".
Once all the factors are bucketed, each factor will be examined for a few things:
- Does it have a sufficient Information Value and/or Gini Coefficient?
- Is the Weight of Evidence monotonic? (Weight of evidence measures how "good" people in this bucket are relative to people in the other buckets)
- Does the factor logically and intuitively rank risk?
The first point talks to whether the factor can distinguish between high and low risk people. The second talks to whether the factor ranks risk in a particular direction. The final point talks to whether the factors are reasonable (e.g. statistically what a person's favourite colour is might be very predictive but there is no logical reason for it. Also there is no reason why people who like red should be better than people who like blue or vice versa)
Now the list of factors have been narrowed down to a much shorter list but a list of factors is still not a model. At this point it is time to build the actual statistical model. There are many techniques to do this but the most common is a stepwise regression (this basically starts with no factors and adds the most predictive excluded factor at each step until a certain accuracy is reached or the additional factors do not add sufficient accuracy to the model as a whole). At this point a lot of analysis is done on the factor selection to make sure that the factors are not too related to each other and that the model is adequately predictive and that the factors included consider every relevant aspect. This is more art than science.
The next step is to decide on the scores. For these models, it is usually specified by 2 parameters: a score for a certain odds of non-payment and a number of point to double (or halve) these odds. There isn't really an industry standard for this so there can be vastly different choices across different institutions but it doesn't actually make a difference to how the model works.
Last step is to assign scores to the individual buckets for the chosen factors. This is pretty simple once all the above steps have been done because it is basically a combination of that bucket's Weight of Evidence and the factor's weighting in the model.
Voila, you now have a scorecard.
The final optional step is using historical data to map the score to a probability that describes the chance of a person with a certain score defaulting on his/her loan over the next 12 months.
Not the point, but ok 
For anyone else interested, read the 4th reply in the thread. 
