Hi.. i'm new in analytic field & trying to learn building Credit risk modeling on my own & facing few difficulties(As I don't know how risk modeling actually gets implemented in Banking  system) After reading few thesis papers & information available on internet i developed some understanding of credit risk modeling(PD) & thus prepared steps that needs to be taken while developing risk modeling which is as below:(Logistic Regression)

1. Definition of Bad:

Bankruptcy process already started or payment due for more than 90 days(basel-2)

2. Sampling

• Sample size needs to be decided to build the model.
• General Formula used={ (Z(alpha/2)sqrt(PDmax(1-PDmax))/Del PD}^2

PD=Max prob of default estimated by experts through analyzing previous years data

Del PD= PD error

• Logistic Regression recommends to use 80% of Good & 20% of Bad thus accordingly we need to modify our data.i.e deleting the goods & increasing the bads.

Data Manipulation

• Missing value treatment
• Outlier detection mainly univariate analysis
• Outliers treatment

-replace all variable having z score > 3 by mean + 3* S.D(Proc Standard)

- replace all variable having z score <- 3 by mean - 3* S.D

-How to decide number of group (Expert judgement?) & how grouping can be done

3. Analysis of Variables choosing statistical model form & calculation of coefficients

• Variables can be chosen through

A. Expert judgements

B.Economic logic, discriminatory power of variables (& other criteria needs to be considered

• Input Selection

-       Cramer’s V,chi sq

-       P-value

• Grouping of attribute values

-       Based on percentile/Attributes (I’m confused with SAS code for the same though I know pctlpts can be used for calculating percentile but do we need to specify each & every variables...Really confused with this...in some paper it was mentioned that grouping should be done based upon WOE...

-       Code the values-apply weight of evidence(WOE)

-       Calculate Information Value

C. Applying forward selection (Mostly used)

4.Validation of Model

-       ROC

-       Ginni coefficient

-       Lorenz curve

-       KS distance

5.Implementing score Card Scaling

-       Score=offset  + factor*ln(odds)Score + PDO(Points to double the odd)

-       Factor=PDO/ln(2)

-       Offset=Score(factor*ln(Odds))

6.Reject Inference

-       Augmentation

-       Extrapolation

If possible please help me out as i don't have any contacts in Analytic field who can guide me..or please suggest me some really good stuff so that i learn read & learn on my own..