Statistically valid scoring for consumer credit applications in banking

Svjetska tržišta bankovnih kredita imaju dugu tradiciju. Razvijena su, postoje i uspješno posluju godinama. U Hrvatskoj je tržište bankovnim kreditima u velikom usponu i još uvijek se razvija. Dubinska analiza podataka nalazi široku primjenu u razlicitim djelatnostima, pa tako i na financijskom tržištu. Jedan od temeljnih zadataka u podrucju bankarskog poslovanja je i razlucivanje dobrih klijenata od loših, odnosno predvidanje vjerojatnosti neplacanja duga vjerovnika. Postupci dubinske analize podataka omogucuju prepoznavanje skrivenih informacija u podacima koje posjeduju financijske institucije i otkrivanja novog znanja. Kreditno bodovanje oslanja se na prepoznavanje rizicne skupine koja je sklona neplacanju. Cilj je predvidjeti i smanjiti stopu neplacanja i delinkventnost klijenata, na osnovu znanja o postojecim klijentima koja se otkrivaju iz povijesnih informacija. U ovom radu iznesene su osnovne postavke i ideje na kojima se zasniva dubinska analiza podataka. Opisani su osnovni koraci u procesu otkrivanja znanja, te metode modeliranja statisticki valjane bodovne kartice. Za svaku pojedinu metodu definirane su njihove specificnosti, svojstva te njihove dobre i loše strane. Konacan cilj bio je primijeniti opisane metode u izradu bodovne kartice na odabranom uzorku i usporediti njihove rezultate. Provedena analiza i rezultati pokazuju da se sve navedene metode mogu koristiti u izradi statisticki valjane bodovne kartice, ali ne pokazuju jednaku sposobnost prepoznavanja dobrih od loših klijenata. Logisticka regresija pokazala je da je njena prediktivna snaga veca u odnosu na ostale metode i kao takva vec je naširoko prihvacena kao najbolja metoda za izradu bodovnih kartica za odobravanje kredita.World market of bank loans has a long tradition; it is highly developed, and successfully operating for years. In Croatia, the market for bank loans is still in development and evolving. Data mining techniques are widely used in various industries, including financial market. One of the fundamental tasks in the field of banking industry is distinguishing good from bad clients, to be more precise predicting the probability of default of debt creditors. Data mining techniques allow recognition of hidden information in data that financial institutions posses and derivation of new knowledge. Credit scoring relies on identifying groups of customers with high risk that are inclined to credit default. The purpose of credit scoring is to predict and reduce default rate and client delinquency based on the knowledge of existing clients that is derived from historical data. This thesis summarises basic assumptions and ideas on which data mining is based. Basic steps in the process of knowledge discovery and statistically valid scorecard modelling methods are described and each method is defined by their characteristics, properties and their pros and cons. The ultimate aim was to apply the methods described in the scorecard development of the selected sample and compare their results. Conducted analysis and experimental results showed that all the methods can be applied when developing statistically valid scorecards, as they all have shown ability to distinguish good from bad clients, but with different scorecard power. The method that has shown to have the highest predictive power is the logistic regression that has already been wide accepted as the best method for application scorecard development