Well, the answer is: Numerous.
In order to make a comment about the methods used for credit scoring, one has to know the idea behind credit scoring. Moreover, the processes involved in building a model for credit scoring also entails knowing the possible methods that can be used. Because some of the methods used has advantages in one area but drawbacks in other areas.
The following part which is taken from the academic article “What is the point of credit scoring?” very well summarizes what can be done and what has been done up to know. Article was written by Loretta Mester who is a vice president and economist in the Research Department of the Philadelphia Fed. She is also the head of the department’s Banking and Financial Markets section.

Scoring Methods
Several statistical methods are used to develop credit scoring systems, including linear probability models, logit models, probit models, and discriminant analysis models. (Saunders discusses these methods.) The first three are standard statistical techniques for estimating the probability of default based on historical data on loan performance and characteristics of the borrower. These techniques differ in that the linear probability model assumes there is a linear relationship between the probability of default and the factors; the logit model assumes that the probability of default is logistically distributed; and the probit model assumes that the probability of default has a (cumulative) normal distribution. Discriminant analysis differs in that instead of estimating a borrower’s probability of default, it divides borrowers into high and low default-risk classes.

Two newer methods beginning to be used in estimating default probabilities include options pricing theory models and neural networks. These methods have the potential to be more useful in developing models for commercial loans, which tend to be more heterogeneous than consumer or mortgage loans, making the traditional statistical methods harder to apply. Options-pricing theory models start with the observation that a borrower’s limited liability is comparable to a put option written on the borrower’s assets, with strike price equal to the value of the debt outstanding. If, in some future period, the value of the borrower’s assets falls below the value of its outstanding debt, the borrower may default. The models infer the probability a firm will default from an estimate of the firm’s asset-price volatility, which is usually based on the observed volatility of the firm’s equity prices (although, as McAllister and Mingo point out, it has not been empirically verified that short run volatility of stock prices is related to volatility of asset values in a predictable way. Saunders discusses other assumptions of the options-pricing approach that are likely to be violated in certain applications.) Saunders reports that KMV Corporation has developed a credit monitoring model based on options-pricing theory.

Neural networks are artificial intelligence algorithms that allow for some learning through experience to discern the relationship between borrower characteristics and the probability of default and to determine which characteristics are most important in predicting default. (See the articles by D.K. Malhotra and coauthors and by Edward Altman and coauthors for further discussion.) This method is more flexible than the standard statistical techniques, since no assumptions have to be made about the functional form of the relationship between characteristics and default probability or about the distributions of the variables or errors of the model, and correlations among the characteristics are accounted for.

Some argue that neural networks show much promise in credit scoring for commercial loans, but others have argued that the approach is more ad hoc than that of standard statistical methods. (The article by Edward Altman and Anthony Saunders discusses the drawbacks.) A study by Edward Altman, Giancarlo Marco, and Franco Varetto analyzed over 1000 healthy, vulnerable, and unsound Italian industrial firms from 1982-92 and found that performance models derived using neural networks and those derived using the more standard statistical techniques yielded about the same degree of accuracy. They concluded that neural networks were not clearly better than the standard methods, but suggested using both types of methods in certain applications, especially complex ones in which the flexibility of neural networks would be particularly valuable.”