Innovation in the Retail Banking Industry: Credit Scoring ... - fep.up… V/V.L... · requirement, a bank will have to segment by credit scores, including application scoring (score

Innovation in the Retail Banking Industry:Credit Scoring Adoption and Consequences on Credit

Availability∗

Marcello Bofondi and Francesca Lotti†

Bank of Italy, Research Department

March 11, 2005

AbstractCredit scoring techniques represent a major innovation aimed at reducing the costsof underwriting processes and improving the accuracy of the estimates of borrowers’default probability. The benefits deriving from the diffusion of credit scoring techniquesare potentially very large, both from the borrowers’ and the banks’ point of view. Weanalyze the patterns of diffusion of this technology among Italian banks. We find thatcredit scoring was first introduced by large banks which are fully able to exploit scaleeconomies. Moreover, a relevant channel of diffusion of this technology is representedby the bank group. We then analyze the effects of credit scoring on the supply ofnew mortgage loans by individual banks, finding that adoption allows them to increasetheir market share. Finally, we examine whether the introduction of automated creditscoring techniques has increased or reduced credit availability in local markets. Theempirical evidence indicates that introduction of credit scoring has a positive andsignificant impact on the overall supply of credit.

Keywords: Innovation, Credit Scoring, Mortgage Loans, Technology Diffusion, RetailBanks.JEL Classification: G21, C41.

∗Bank of Italy, Research Department, via Nazionale 91, 00184 Rome, Italy. Email: [email protected]; [email protected]. We wish to thank Giorgio Gobbi and FabioPanetta for their valuable comments, Jean Marie Bouroche and Alessandra Gabrielli of CRIF Decision Solu-tion for helping us in understanding how credit scoring is designed and for data provision. A special thank tothose colleagues at the Regional Research Units who helped us with the survey. Cinzia Chini and M.CristinaFabbri provided excellent research assistance. The usual disclaimers apply. The views expressed here do notnecessarily reflect those of the Bank of Italy.† Corresponding author.

1

1 Introduction

In this paper we study the diffusion of credit scoring technologies among Italian banks and

its consequences on credit availability to households. Automated credit scoring techniques

represent a major innovation aimed at reducing the costs of underwriting processes and

improving the accuracy of the estimates of borrowers’ default probability. Therefore, it

must be considered a new production process (Frame and White, 2004).

The first application of credit scoring techniques took place in the US, where they were

primarily used for credit cards loans. In the last decade credit scoring has become very

common in the mortgage lending industry. Two US Congress chartered companies (Fannie

Mae and Freddie Mac) contributed significantly to the diffusion of this technology. Fannie

Mae and Freddie Mac where chartered to stabilize the mortgage market and expand op-

portunities for home ownership. They accomplished their mission by creating a secondary

market for mortgage loans. In order to reduce the informational asymmetries between loans

originators and subscribers of the securities backed by the mortgage loans, these two com-

panies require the originator to use automated underwriting technologies based on credit

scoring techniques. Their experience in the mortgage loans market and the availability of a

large number of data allowed them to develop their own automated underwriting software,

that is currently sold to the originators. Nowadays, in the US mortgage loans market, the

creditworthiness of virtually all the new borrowers is evaluated by credit scoring techniques.

Starting from the mid nineties, credit analysts argued that this technology could have

been profitably applied also to small business lending since the personal credit history of small

business owners is highly predictive of the loan repayment prospects of the business. This

implied the implementation of automated underwriting in a market that was traditionally

characterized by a pervasive use of soft information. The patterns of diffusion of credit

scoring in the small business loans market and the effects on availability, price and risk

of small business credit have been studied by some recent contributions (Akhavein, Frame

and White, 2004; Berger, Frame and Miller, 2004). The main findings of these studies

are that the first adopters were mainly large banking organizations characterized by fewer

separately chartered banks, but more branches. Moreover, credit scoring is associated with

expanded quantities, higher average prices, and greater risk levels for small business credit.

These findings are consistent with a net increase in lending to relatively risky “marginal

2

borrowers” who would not otherwise receive credit, but who pay relatively high prices when

they are founded.

In Italy credit scoring diffusion is still at an early stage. This is due to several reasons.

First, the lack of comprehensive Credit Bureaus providing potential borrowers’ credit his-

tories, can complicate the construction of a standardized credit scoring mechanisms. The

largest Italian Credit Bureau i.e. the Italian Credit Register (C.R.), files information about

all bad loans and performing loans above 75,000 euro. This implies that most of information

concerning consumer credit is not reported. Therefore a complete borrower’s credit history

cannot be tracked. In order to fill this gap, some private companies have started to gather

information about those loans not filed in the C.R.. Nevertheless these private credit bureau

are not as detailed as those maintained in the US. Second, in Italy the secondary market for

loans was thin until the late 90’s, and consequently it provided little incentives for the adop-

tion of standardized credit evaluation methods. Finally, the relationships between banks

and their customers, which in Italy rely heavily on soft information, make the adoption of

automated credit scoring techniques more difficult, in particular for small business lending.

The New Basel Capital Accord will constitute a great incentive for the diffusion of credit

scoring. In order to implement the so called Internal Rating-Based approach, banks will be

required to segment their retail portfolio according to borrowers’ credit risk. As a minimum

requirement, a bank will have to segment by credit scores, including application scoring

(score based on full information in a credit application).1

The potential benefits from the diffusion of credit scoring techniques are very large. From

the borrowers’ point of view, automated underwriting simplifies the loans application pro-

cess. Borrowers are asked a limited and standardized set of questions about their financial

and socio-economic conditions and are provided with a quick answer, allowing them to re-

duce the search costs that they would have incurred under the traditional loan evaluation

processes. Further, banks can significantly reduce the marginal cost of creditworthiness eval-

uation, allowing them to assess a larger number of potential borrowers. Credit scoring might

be used as well during renegotiations, permitting the bank to redefine contracts’ conditions.

As long as these techniques are more reliable than traditional judgmental methods, their

implementation could also reduce credit risk. Finally, from a social welfare perspective, the

1See Basel Committee on Banking Supervision (2001) and Altman (2002).

3

use of credit scoring can facilitate the creation of a secondary market for loans, increasing

credit availability by allowing banks to invest in liquid assets and to improve risk manage-

ment.2 Nevertheless there still are some potential shortcomings. First, the possibility of

screening a large number of borrowers may induce banks not to expand their credit supply,

but to select more carefully their borrowers. As a consequence we might observe a decrease

in credit availability. Second, if the automated tool is not flexible enough, there is risk of

systematic exclusion of certain categories of new borrowers (Bridges and Disney, 2001), with

a negative impact on social welfare. This has been a major concern in the US mortgage

loans market in the last few years.

Our analysis is based on a database obtained from a survey that we conducted among

104 banks that had originally reported to the Supervision Department of the Bank of Italy

the usage credit scoring techniques. We received 45 answers, 43 of them confirmed their

adoption of credit scoring and provided further information, such as the date of adoption

and the market segment to which credit scoring is applied. Complementing this data with the

information coming from the Banking Supervision Reports at the Bank of Italy, we analyze

the pattern of diffusion of credit scoring among Italian banks starting from 1993. We find

that the adoption wave started in 1998, that first adopters were mainly large commercial

banks and that a main channel of diffusion was the banking group. We also estimate a non

parametric hazard model finding that the curvature of the cumulative hazard curve is very

steep, suggesting that we still are in the first stages of the diffusion process. Afterwards we

focus our attention on the consequences of credit scoring adoption on the market of mortgage

loans, using a data set of 18,245 observations referring to 103 provinces and 240 banks over

the period 1999-2002. Our main findings are that the adoption of credit scoring implies a

faster growth rate of new mortgage loans for the adopters and that this expansion is mainly

due to the possibility of financing customers that otherwise would have not been financed.

2 An overview on the CS methodology

Credit scoring is a technique used by financial intermediaries for screening applicants. This

procedure was first adopted during the fifties, mainly by finance houses and mail order

2Whether securitzation may help to lower interest rates is still an open issue (Heuson et al., 2000).

4

firms. At that time, decision were made judgementally by credit analyst: creditworthiness

evaluations largely depended on the rules of financial houses and on the experience and

knowledge of their clerks. In the sixties, when the number of applicants for credit cards

increased, an automated system was necessary. The first consultant firm was founded in

1957 in San Francisco (Bill Fair & Earl Isaac); since then, the automated credit scoring

procedure spread quickly and became a standard for risk evaluation for consumer lending,

for mortgage lending and for business lending.

Credit scoring procedures are based on statistical methods like discriminant analysis,

logistic and probit regression; in more recent developments, neural networks, genetic algo-

rithms and linear programming are used. The common idea behind these different method-

ologies, is that there exist (at least) two populations of potential borrowers, good and bad

types (with a lower and a higher average default probability, respectively). From a statis-

tical point of view, it is a problem of correct classification, i.e. to design a procedure able

to allocate a new observation into one of the two populations, minimizing a given objective

function (typically is the cost of misallocation). Accordingly, individuals are separated on

the basis of some observed characteristics belonging both to their socio-economic background

- age, income, gender, nationality, family size, etc. - and to its credit history - the number

of credit cards, how much did the applicant borrowed, if it ever delayed a repayment and so

on.3

As an example, a problem of classification of an individual in two population is sketched.

Let’s define f1 (x) and f2 (x) the probability density functions associated with the random

vector X - the observable characteristics - for the population φ1 (good types) and φ2 (bad

types). If we denote the sample space with Ω, a potential borrower with a profile must be

assigned to either φ1 or φ2. Suppose then to split the sample space into two regions, S1 and

S2, exhaustive and mutually exclusive, and to classify as φ1 (φ2) those observations for which

x ∈ S1 (S2). Since the probability distributions f1 (x) and f2 (x) are often overlapping, the

assignment process described above, is potentially subject to misclassification errors. To give

an example, observations belonging to the S1 (S2) region can be erroneously be assigned to

population φ1 (φ2) while coming from φ2 (φ1), as represented Figure 1 with the light (dark)

gray area. Each of these possible misallocation has a different cost: accordingly, one criterion

3In the US, the Equal Credit Opportunity Act (ECOA) prevents creditors from discriminating potentialborrowers according to some characteristics, such as race, sex, marital status, nationality, and so on.

5

to identify the boundary of the S regions could be the minimization of the expected cost of

misallocation. The statistical models of classification are usually based on historical data

and this rises the issue of sample attrition.4 Since the aim of all these procedures is to

give an estimate of the default, the result can be severely biased. The selection mechanism

works as follows: given a potential pool of borrowers, subdivided in good and bad types,

we can observe creditworthiness performance only of those whose application was accepted;

thus a classification process which does not take into account the source of attrition is

based only on one part of the population. In this way, an applicant’s profile is compared

against successful recipients, therefore conditioning on the sampling rule, i.e. acceptance5

Accordingly, the predictor of the default rate must take into account this rule. Besides

these problems related to the statistical significance of the sample of potential borrowers,

there is a crucial issue concerning social welfare. It may happen that a new applicant

exhibits characteristics that are “infrequent”, in the sense that they are at the boundaries

of the sample space Ω (or sometimes, even outside), and to be rejected just because similar

individuals have no credit history. From a technical point of view, these statistical models

need a population of individuals to extract the sample from: for a single bank that needs

to use these models, the population is the actual pool of its borrowers. Based on this

information, the statistical tool is calibrated: in this sense, scoring mechanism belonging

to different banks are independent. When internal information is inadequately organized or

incomplete, a “start-up” model is implemented6 this is based on a more general representative

sample (for type of loan and geographically) of more vast and generic pool of potential

borrowers.

3 Italian Banks and Technology Adoption

Recently, the Bank of Italy conducted a survey among Italian banks (excluding Cooperative

Credit Banks) about the possible consequences of the New Basel Capital Accord. Nearly

4See Greene, 1998.5In probabilistic terms, if one doesn’t take into account the sample attrition problem, the estimate

will be Prob (Default|Acceptance), which is different from Prob (Default), since Prob (Default) andProb (Acceptance) are not independent.

6In these cases, the supplier of the credit scoring technique is very likely to be an external specializedfirm.

6

one hundred banks declared to have adopted credit scoring techniques. By means of a direct

contact with those banks, we obtained further information about credit scoring adoption7,

such as the date of adoption and the field of application. We received 45 answers: 43 of them

were complete. Then we merged this data with bank level information from the Banking

Supervision Reports at the Bank of Italy, both on the adopters and on the non adopters. The

first adoption occurred in 1989 by a subsidiary of a foreign bank. The true adoption wave

occurred from 1998 onward, as emerges from Table 1. Undoubtedly, the time to adoption is

very likely to be affected, especially at the early stages, by banks’ individual characteristics,

like size, number of branches, market share in the field of application of the credit scoring or

other non observable strategies. Concerning bank size, we would expect larger institutions

to be among the firsts to adopt credit scoring.8 Conversely, smaller banks which tend to

rely more on soft information are expected to be later adopters. A channel through which

diffusion takes place is the banking group: once the head of a group adopts automate credit

scoring techniques, it imposes the same standards to the other banks belonging to the group.

This is shown in Table 1 (column 6): in 2002, 77% of the adopters in our sample were part

of a group. The weight of adopter banks increased over time: at the end of 2002 they

accounted for the 40% of outstanding mortgage loans (from 18% in 1999) and for the 51%

of outstanding consumer credit (from 26% in 1999). The large proportion of debt extended

by adopters banks is due to the fact that, even if the number of adopters still seems small,

they represent the larger Italian banks. Columns 4 and 5 of Table 1 reports the number of

adopters from 1993 on, categorized by bank types.9 The pattern of diffusion is similar across

bank types, even though early adopters were primarily Commercial Banks. In 2002, 67%

of adopters were Commercial Banks and 33% Cooperative Banks. We also classify bank in

term of size, according to the amount of total loans: top banks (over 45 billions of euro),

big banks (from 20 to 45 billions of euro), medium sized banks (from 7 to 20 billions of

euro), small banks (from 1 to 7 billions of euro) and minor banks (less than 1 billion of

euro). In Figure 3, adopters are classified according to their size: top and medium sized

banks are the pioneers in using automated credit scoring techniques. With the additional

information gathered directly through the banks, we could identify the field of application

7The questionnaire employed is in Italian, and it is available from the authors upon request.8Larger banks can benefit from scale economies, moreover, they are likely to have a more codified customer

database, necessary to implement a customized credit scoring mechanism.9Cooperative Credit Banks are not taken into account since they are not part of our sample.

7

of the automated credit scoring techniques. In Figures 4-5-6, the patterns of adoption of

credit scoring for consumer credit (credit cards and personal loans), mortgage loans and

small business lending are reported. The pattern of diffusion resembles the one occurred in

the US: credit scoring was firstly implemented to assess the creditworthiness of borrowers

applying for credit cards and consumer credit. The strong links between the probability of

default and the personal credit histories of the borrowers make this type of loans particularly

suitable for the use of automated underwriting. Lately banks started to apply credit scoring

to mortgage loans as well. The use of this new technology in the small business credit market

is still not very disseminated, indicating how soft information can still play a fundamental

role especially in small business lending practices. In order to better examine the speed and

of adoption and the stage of the diffusion process, we estimate a non parametric hazard

model, subdividing the period from 1989 to 2002 (in days) and evaluating, for every bank

at every time interval, what the likelihood of adoption is, conditional on the fact that credit

scoring was not used before.

Formally, the hazard rate can be defined as:

λ (t) = lim∆t→0

Pr (t ≤ T ≤ t + ∆t|t ≤ T )

∆t=

f (t)

1− F (t)(1)

where f (t) is the density function whose cumulative density function is F (t). For each

time t, the hazard rate λ (t) is the probability of adopting the new technology in the period

t + ∆t, given that the bank did not adopt credit scoring at time t. For our purpose, we use

the integrated (or cumulated) hazard function, defined as:

Λ (t) =

∫ t

0

λ (t) dt (2)

which can be consistently estimated through the Nelson-Aalen estimator (Aalen, 1978 and

Nelson, 1972), defined as:

H (t) =k∑

j=1

(δj

nj

)(3)

where δj represents the number of banks which adopted credit scoring in the jth time

interval, nj is the number of non adopters at the beginning of the jth time interval and k is

8

the number of time intervals considered. In Figure 8, the Nelson-Aalen cumulative hazard

function is reported.10 The starting point, i.e. the time in which the first adoption occurred,

is August 1989: up to that time we observe a nearly flat line. Only after 3 thousands days

since the first adoption (i.e. more than eight years), this innovation started to spread, that

means that the true process of diffusion took place from 1998 to now. When the adoption

process is concluded, we would expect to observe a complete hazard curve S−shaped: in the

period from 1998 to 2002 the curvature of the hazard curve is very steep, suggesting that we

still are in the first stages of adoption (Geroski 2000).

4 Consequences on credit availability: empirical evi-

dence

We focus our empirical analysis on the mortgage loans market. The US experience suggests

that this is a market where the implementation of credit scoring techniques may have sub-

stantial consequences. The market for mortgage loans in Italy expanded rapidly in the past

four years: in 2002, 37 billions of euro of new mortgage loans were extended, 33% more than

in 1999. In the same period, also the implementation of credit scoring in the assessment of

credit worthiness of borrowers applying for a new mortgage loan expanded rapidly: accord-

ing to the results of our survey, in 1999 in every Italian province, on average, there were at

least 4.7 banks using credit scoring, in 2002 they were 10.9 (Table 2).

We conduct our empirical analysis on a sample referring to the period 1999-2002 and

to all Italian banks (excluding Cooperative Credit Banks) operating in the mortgage loans

market and for which data were available. Our sample indicates a significant growth of the

diffusion of credit scoring techniques. In 1999 the share of new mortgage loans extended by

adopting banks was equal to 29.5%, that became 31.5% in 2002. In the same period, the

number of provinces where more than 40% of new mortgage loans were extended by adopters

rose from 17% to 28%. In 1999 there where 40 provinces with a share smaller than 20%, in

2002 only 15.

The adoption of automated credit scoring mechanisms can have effects on credit avail-

ability both at the market and at the bank level. For this purpose we estimate two empirical

10Taking into account that the sample is right censored; time is expressed in days.

9

models: the first aims at taking into account its effect on the expansion of new mortgage

loans extended by the single banks, while the second at assessing the impact of credit scoring

on market size.

4.1 Consequences on banks’ mortgage loans expansion

We analyze the impact of credit scoring adoption on new mortgage loans extended by bank

i in province j at time t (LNMORTi,j,t). We assume, as shown in equation (1), that that

(LNMORTi,j,t) depends on whether the bank has adopted credit scoring at time t − 1

(DCSi,j,t−1), on banks characteristics and on two dummy variables one for province j and

one for time t.

LNMORTi,j,t = f(DCSi,j,t−1, SIZEi,j,t−1, CAPITALi,j,t−1, LNCOSTSi,j,t−1, (4)

BRANCHESi,j,t−1, BRMKTi,j,t−1, DGRi,j,t−1, DPROVj, DTIMEt)

Bank’s i size (SIZEi,j,t−1) is measured in terms of total assets, its capital (CAPITALi,j,t−1)

defined as the ratio of equity capital to total assets. Cost efficiency is accounted for by the ra-

tio between operating expenses and total assets (LNCOSTSi,j,t−1); we also add network size

as a regressor, measured with the total number of branches of bank i (BRANCHESi,j,t−1).

The presence in the market is captured by the variable BRMKTi,j,t−1, which represents the

number of branches of bank i in province j. Finally, a dummy variable indicating whether

bank i belongs to a group is included. All bank-level variables refer to period t − 1 to

avoid possible endogeneity problems. The province dummy variable controls for all pos-

sible province fixed effects (i.e. different demand, market structure and overall economic

conditions), while the time dummies captures common time shocks. We use a similar spec-

ification to answer a different question: does mortgage loans supply depend upon the time

since credit scoring adoption? We expect the supply of mortgage loans to depend positively

on time since adoption, but at declining rates over time. For this reason, we substitute the

dummy variable DCSi,j,t−1 with the numbers of years since adoption (TIMEi,j,t−1) and a

squared term (TIMESQi,j,t−1). We estimate the different specifications using a panel data

base of 18,245 observations referring to 103 provinces and 240 banks over the period 1999-

10

2002. A full description of the variables can be found in the appendix, while descriptive

statistics are shown in Table 3. The estimation method is OLS regression with province

fixed-effects.11 The results are reported in Table 5. Column 1 displays the estimates for

equation (4): the coefficient of the credit scoring adoption dummy variable is positive and

significant (the coefficient is 0.303). Assuming that the expansion of new mortgage loans de-

pends linearly on time since adoption (column 2), we obtain a positive coefficient, implying

one more year of use of credit scoring increases by 0.5% the extension of mortgage loans. Our

hypothesis about a possible non linearity of this relationship is confirmed by the coefficient

of the squared term introduced in equation (4), TIMESQ, which is negative and significant.

The value of the coefficient (−0.022) implies that the growth rate of new mortgage loans is

increasing during the first 4 years since adoption and declining afterwards.

4.2 Consequences on credit availability at the market level

The empirical evidence provided above is consistent with the hypothesis that the adoption

of credit scoring implies a faster growth rate of new mortgage loans. This may happen

in two possible ways: either adopters increase their market share at the expenses of their

competitors, or credit scoring allows to finance borrowers that otherwise would not have been

financed. In the former case, adoption would have consequences only on adopter banks. In

the latter, the whole market would result enlarged. In order to identify which of the two

effects is the prevailing one, we model new mortgage loans extended in province j at time t

(LNMORTMKTj,t) as a function of province level variables and information about credit

scoring diffusion, as described in equation (5).

LNMORTMKTj,t = f(WEIGHTj,t−1, HERFj,t−1, RATESj,t−1, (5)

GDPj,t−1, REALESTj,t−1, DTIMEt)

Equation (5) measures credit scoring diffusion as the ratio between the number of adopt-

ing banks on the total number of active banks in province j at time t− 1 (WEIGHTj,t−1).

We control for province specific effects with: the degree of competition measured by the

11We did not include in the model bank-fixed effect which would have captured all the variations in thecharacteristics of interest, since variance over time was relatively small.

11

Herfindahl Index - computed on loans - (HERFj,t−1); the average loan rate (RATESj,t−1)

as a proxy for the prevailing risk conditions; the log of real per capita GDP (GDPj,t−1), as an

indicator of the local economic conditions and the real estate average price (REALESTj,t−1),

to account for changes in the demand for real estates. All province-level variables refer to

period t−1 in order to avoid possible endogeneity problems. Subsequently we exploit the in-

formation embedded in the variable WEIGHT , substituting it with the number of adopting

banks (SCORINGj,t−1) and the total number of banks in the market (NBANKSj,t−1).12

In an alternative specification we replaced the NBANKS variable with the Herfindahl In-

dex. We estimate the three specifications described above using a panel database of 421

observations referring to 103 Italian provinces over the period 1999-2002. A full description

of the variables can be found in the appendix, while descriptive statistics are shown in Table

4. We estimate a simple OLS regression since the variance over time is very small compared

to the overall variance.13 The results are reported in Table 6. The impact of the share of

adopters on the availability of new mortgage loans is positive and highly significant. The

coefficient is 8.354, implying a semielasticity, computed at the mean value, of 0.25 (column

1). Therefore, the hypothesis of credit market expansion as a consequence of credit scoring

adoption is corroborated by empirical evidence. The advantages of credit scoring, namely

the possibility of assessing creditworthiness of potential borrowers quickly and at relatively

lower marginal cost, allow banks to increase credit supply. This finding is robust to different

measures, like the number of adopting bank tout court (given the number of active banks

in the market), which has a positive and significant impact on new mortgage loans. The

coefficient, equal to 0.066, implies a semielasticity, computed at the mean value, of 0.30

(column 2). The coefficient of the Herfindahl Index, even if is not statistically significant,

indicates that a higher degree of competition can increase the availability of new mortgage

loans. The province controls are all highly significant and consistent across specifications.

12In this specification we dropped the HERF variable to avoid collinearity problems with the NBANKSvariable.

13Actually, we estimated also a fixed effects and random effects model, but the within variance was toosmall to justify the use of this estimation technique. For this reason we preferred a simple OLS estimationwith time dummies.

12

5 Conclusions

In this paper we analyzed the pattern of diffusion of credit scoring techniques among Italian

banks. We estimate a non parametric model finding that, though the early adopters started

to implement this technology in the late 80’s, the true adoption wave started in 1998. In

the period from 1998 to 2002 the curvature of the hazard function is very steep, suggesting

that we still are in the first stages of adoption. The patterns of diffusion are significantly

different across dimensional classes: larger banks turns out to be among the first adopters,

and this is consistent with the fact that those banks can benefit from scale economies more

than their smaller counterparts. Moreover, a relevant channel of diffusion of this technology

is represented by the bank group. From a credit availability point of view, we address two

main issues. The first is related to the effects of credit scoring adoption on the supply of

new mortgage loans by single banks. The results of our econometric analysis shows that

adopter banks actually widen their supply of new mortgage loans more than non adopters.

Time since adoption is also important, in particular in the first years: after four years (on

average) the propulsive effect of credit scoring ceases to be so relevant. The second issue

concerns whether the introduction of automated credit scoring techniques have increased

or reduced overall credit availability in local markets. The empirical evidence indicates

that introduction of credit scoring has a positive and significant impact on market size: a

growth in the number of banks using this technique increases new mortgage loans in a given

market. Further issues concerned with the diffusion of credit scoring are the consequences

of the adoption of this technology on banks’ risk-taking and risk-management, we intend to

address them in future research.

13

References

[1] Aalen, O. O., 1978, “Non Parametric Inference for a Family of Counting Processes”,Annals of Statistics, 6, 701-726.

[2] Akhavein J., W. S. Frame and L. J. White, 2004, “The Diffusion of Financial Inno-vations: Adoption of Small Business Credit Scoring by Large Banking Organizations”,forthcoming in Journal of Business.

[3] Altman, E. I., 2002, “Revisiting Credit Scoring Models in a Basel 2 Environment”,mimeo Stern School of Business, NYU.

[4] Basel Committee on Banking Supervision, 2001, “Consultative Document. The NewBasel Capital Accord”, Bank for International Settlements, Basel.

[5] Berger, A. N., W. S. Frame and N. H. Miller, 2004, “Credit Scoring and the Availability,Price, and Risk of Small Business Credit”, forthcoming in Journal of Money, Credit andBanking.

[6] Bridges S. and R. Disney, 2001, “Modelling Consumer Credit and Default: The ResearchAgenda”, Experian Centre for Economic Modelling.

[7] Frame, W. S. and L. J. White, 2004, “Empirical Studies of Financial Innovation: Lotsof talk, Little Action?”, Journal of Economic Literature, 42, 116-144.

[8] Geroski P.A. (2000), “Models of Technology Diffusion”, Research Policy, 29, 603-625.

[9] Greene W. (1998), “Sample Selection in Credit-Scoring Models”, Japan and the WorldEconomy, 10(3), 299-316.

[10] Heuson, A., Passmore W. and Sparks R., 2000, “Credit Scoring and Mortgage Securi-tization. Implications for Mortgage Rates and Credit Availability”, mimeo.

[11] Nelson, W., 1972, “Theory and Applications of Hazard Plotting for Censored FailureData”, Technometrics, 14, 945-965.

14

Figure 1: Example of two populations with overlapping distributions (univariate case).

φ φ

15

Fig

ure

2:A

dop

ters

ofC

redit

Sco

ring

for

Loa

ns:

clas

sifica

tion

acco

rdin

gto

ban

ksi

ze.

16

Fig

ure

3:A

dop

ters

ofC

redit

Sco

ring

for

Con

sum

erC

redit

:cl

assi

fica

tion

acco

rdin

gto

ban

ksi

ze.

17

Fig

ure

4:A

dop

ters

ofC

redit

Sco

ring

for

Mor

tgag

eLoa

ns:

clas

sifica

tion

acco

rdin

gto

ban

ksi

ze.

18

Fig

ure

5:A

dop

ters

ofC

redit

Sco

ring

for

Sm

allB

usi

nes

sLen

din

g:cl

assi

fica

tion

acco

rdin

gto

ban

ksi

ze.

19

Figure 6: Nelson-Aalen cumulative hazard function: automated credit scoring adoption(starting point 1989, censored at December 2002). Time expressed in days.

20

Table 1: Number of banks which adopted credit scoring techniques and year of adoption.

Cumulated Number of AdoptersYear Number of

Adopters Total Commercial Cooperative Banks belongingBanks Banks to a group

1993 3 3 2 1 31994 1 4 3 1 41995 1 5 4 1 51996 1 6 5 1 61997 2 8 6 2 81998 4 12 9 3 111999 8 20 14 6 172000 14 34 23 11 282001 5 39 25 14 312002 4 43 29 14 33

Table 2: Number of banks using credit scoring for mortgage loans by province.

Minimum Maximum Average

1999 2 11 4.7

2000 4 16 7.3

2001 3 17 8.8

2002 5 23 10.9

21

Table 3: Descriptive statistics at the bank-province level.

Variable Mean N Max Min Std. Dev.

LNMORT -0.57 18728 7.18 -13.82 2.36

DCS 0.1 18728 1 0 0.3

TIME 0.25 18728 8 0 0.95

TIMESQ 0.96 18728 64 0 5.57

SIZE 8.77 18681 12.26 1.61 1.81

CAPITAL 6.01 18466 9.53 0 1.75

COSTS 0.05 18448 0.69 0 0.03

BRANCHES 306.15 18728 2235 0 371.71

BRMKT 5.1 18728 440 0 16.28

DGR 0.83 18728 1 0 0.38

HERF 0.09 18728 0.24 0.04 0.04

RATES 7.17 18728 11.27 4.16 1.22

GDP 8.93 18728 11.47 7.01 0.88

REALEST 7.38 18728 8.19 6.71 0.3

Table 4: Descriptive statistics at the province level.

Variable Mean N Max Min Std. Dev.

LNMORTMKT 4.92 412 8.61 1.9 1.09

WEIGHT 0.03 412 0.1 0 0.02

SCORING 4.55 412 17 0 3.47

NBANKS 136.18 412 393 61 45.8

HERF 0.09 412 0.24 0.03 0.04

RATES 7.35 412 11.27 4.16 1.23

GDP 8.69 412 11.47 7.01 0.77

REALEST 7.31 412 8.19 6.71 0.28

22

Table 5: Consequences of the adoption of credit scoring on mortgage loans supply at thebank-province level. Ordinary least square estimation with province fixed effects. Dependentvariable: logarithm of new mortgage loans by bank and province. Standard errors in brackets.Statistically different from zero, respectively at: *** 99%, ** 95% and * 90% significancelevel. Time and province dummies not reported for brevity.

(1) (2) (3)

DCS 0.303*** - -(0.046)

TIME - 0.054*** 0.173***(0.013) (0.036)

TIMESQ - - -0.022***(0.006)

SIZE 0.549*** 0.563*** 0.546***(0.031) (0.031) (0.032)

CAPITAL -0.508*** -0.520*** -0.503***(0.029) (0.030) (0.030)

COSTS 6.326*** 6.406*** 6.314***(0.618) (0.622) (0.622)

BRANCHES 0.002*** 0.002*** 0.002***(0.000) (0.000) (0.000)

BRMKT 0.054*** 0.054*** 0.054***(0.004) (0.004) (0.004)

DGR -0.632*** -0.617*** -0.648***(0.050) (0.051) (0.052)

CONST -2.659*** -2.709*** -2.643***(0.217) (0.218) (0.216)

N. obs. 18,245 18,245 18,245F-test 45.73*** 44.86*** 45.21***R-squared 0.313 0.313 0.313

23

Table 6: Consequences of the adoption of credit scoring on mortgage loans availability atthe province level. Ordinary least square estimation. Dependent variable: logarithm of newmortgage loans by province. Standard errors in brackets. Statistically different from zero,respectively at: *** 99%, ** 95% and * 90% significance level. Time dummies not reportedfor brevity.

(1) (2) (3)

WEIGHT 8.354*** - -(2.370)

SCORING - 0.066*** 0.058***(0.017) (0.017)

NBANKS - -0.003*** -(0.001)

HERF -0.400 - -0.272(0.499) (0.491)

RATES -0.137*** -0.136*** -0.143***(0.021) (0.022) (0.022)

GDP 1.024*** 1.079*** 0.938***(0.033) (0.064) (0.040)

REALEST 0.703*** 0.703*** 0.664***(0.117) (0.119) (0.119)

CONST -8.001*** -8.680*** -7.431***(0.791) (0.938) (0.829)

N. obs. 412 412 412F-test 413.53*** 455.29*** 397.46***R-squared 0.875 0.877 0.875

24

Appendix

Tab

leA

1-D

ata

Des

crip

tion

.A

llth

eva

riab

lesar

eye

arly

and

refe

rto

the

per

iod

from

Jan

uar

y19

99to

Dec

ember

2002

.

LN

MO

RT

i,j,

tT

he

log

ofth

enew

mor

tgag

elo

ans

exte

nded

by

ban

ki

inpro

vin

cej

atti

me

t.

LN

MO

RT

MK

Tj,

tT

he

log

ofth

enew

mor

tgag

elo

ans

exte

nded

inpro

vin

cej

atti

me

t.

TIM

Ei,t−

1N

um

ber

ofye

ars

since

ban

ki

adop

ted

cred

itsc

orin

gfo

rm

ortg

age

loan

sat

tim

et−

1.

TIM

ESQ

i,t−

1T

he

squar

eof

the

num

ber

ofye

ars

since

ban

ki

adop

ted

cred

itsc

orin

gfo

rm

ortg

age

loan

sat

tim

et−

1.

WE

IG

HT

j,t−

1R

atio

bet

wee

nth

enum

ber

ofad

opti

ng

ban

ks

onth

eto

talnum

ber

ofac

tive

ban

ks

inpro

vin

cej

atti

me

t−

1.

DC

Si,t−

1D

um

my

vari

able

that

take

sva

lue

1if

ban

ki

had

cred

itsc

orin

gat

tim

et−

1an

d0

other

wis

e.

SC

OR

IN

Gi,t−

1N

um

ber

ofban

ks

inpro

vin

cej

usi

ng

cred

itsc

orin

gat

tim

et−

1.

SIZ

Ei,t−

1Log

ofto

talas

sets

ofban

ki

atti

me

t−

1.

CA

PIT

AL

i,t−

1Log

ofban

ki

capit

alat

tim

et−

1.

CO

ST

Si,t−

1R

atio

bet

wee

nop

erat

ing

expen

ses

and

tota

las

sets

ofban

ki

atti

me

t−

1.

NB

AN

KS

j,t−

1N

um

ber

ofban

ks

inpro

vin

cej

atti

me

t−

1.

BR

AN

CH

ES

i,t−

1N

um

ber

ofbra

nch

esof

ban

ki

atti

me

t−

1.

BR

MK

Ti,j,

t−1

Num

ber

ofbra

nch

esof

ban

ki

atti

me

t−

1in

pro

vin

cej.

HE

RF

j,t−

1H

erfindah

lin

dex

com

pute

don

loan

sin

pro

vin

cej

atti

me

t−

1.

RA

TE

Sj,

t−1

Ave

rage

inte

rest

rate

son

loan

sin

pro

vin

cej

atti

me

t−

1.

GD

Pj,

t−1

Log

ofth

eper

capit

are

alG

DP

inpro

vin

cej

atti

me

t−

1.

RE

AL

EST

j,t−

1A

vera

gere

ales

tate

pri

cein

pro

vin

cej

atti

me

t−

1.

DG

Ri,t−

1D

um

my

vari

able

that

take

sva

lue

1if

ban

ki

was

par

tof

agr

oup

atti

me

t−

1an

d0

other

wis

e.

25

Innovation in the Retail Banking Industry: Credit Scoring ... - fep.up… V/V.L... · requirement, a bank will have to segment by credit scores, including application scoring (score

Documents