Financial incentives and loan officer behavior: multitasking and allocation of effort under an incomplete contract

Financial Incentives and Loan Officer Behavior:
Multitasking and Allocation of Effort under an Incomplete Contract
Patrick Behr
EBAPE, Getulio Vargas Foundation
Alejandro Drexler
Federal Reserve Bank of Chicago
Reint Gropp
IWH, University of Magdeburg and SAFE
Andre Guettler‡
University of Ulm
This version: November 13, 2015
Abstract
We investigate the implications of providing loan officers with a compensation structure that rewards loan
volume and penalizes poor performance. Using a unique data set provided by a large international
commercial bank, we examine the three main activities that loan officers perform: monitoring, origination,
and screening. We find that when loan officers are at risk of losing their bonus, they increase monitoring
and origination, but not screening effort. On the other hand, having lost a bonus in the previous period does
not entail higher effort. We document unintended consequences of the incentive contract showing the
incompleteness of such contracts.
JEL Classifications: G21, J33
Keywords: Loan officer, incentives, monitoring, screening, loan origination

We would like to thank Phil Dybvig, Mrdjan Mladjan, Lars Norden, Klaus Schaeck, Jerome Taillard, participants at
the annual meetings of the American Finance Association 2015, the European Finance Association 2015 and the Swiss
Society for Financial Market Research 2015, and seminar participants of the Annual Meeting on Risk, Financial
Stability and Banking of the Brazilian Central Bank, Bank of Canada, Brazilian School of Public and Business
Administration, Federal Reserve Bank San Francisco, Frankfurt School of Finance and Management, Fudan
University, and World Bank for their comments. Part of this research was completed while Gropp was visiting the
University of Amsterdam, the hospitality of which is greatly appreciated. A previous version of the paper was entitled
"Financial Incentives and Loan Officer Behavior". The views expressed here do not reflect official positions of the
Federal Reserve Bank of Chicago.

Brazilian School of Public and Business Administration (EBAPE), Getulio Vargas Foundation (FGV), Praia de
Botafogo 190, 22253-900 Rio de Janeiro, Brazil, Email: patrick.behr@fgv.br.

Federal Reserve Bank of Chicago, 230 South LaSalle St., Chicago, IL 60604, United States, Email:
alejandro.h.drexler@chi.frb.org.

Institute for Economic Research Halle (IWH), University of Magdeburg and SAFE. Email: reint.gropp@iwh-
halle.de.
‡
University of Ulm, Institute of Strategic Management and Finance, Helmholtzstraße 22, 89081 Ulm, Germany,
Email: andre.guettler@uni-ulm.de.

1
1. Introduction
While most research on bank compensation focuses on equity-linked incentives for high-level managers,
there seems to be a consensus that distorted financial incentives for lower-level employees, such as loan
officers and loan originators, was one of the factors at the root of the 2007–2008 financial crisis. The role
of loan officers’ behavior in the crisis opened a controversial public debate that has already caused
important changes in the regulatory framework.1
The debate has also increased academics’ and
practitioners’ interest in exploring the implications of incentive-based compensation for financial
institutions.
In this paper we use detailed data on a non-linear compensation structure of loan officers at a large
international bank to study how financial incentives affect their behavior. In our sample period, the bank
used an incentive-based compensation structure, in which loan officers received a monthly cash bonus
proportional to their lending volume besides their fixed salary. However, the bonus was not paid in months
when the value-weighted non-performing loans in a loan officer’s portfolio exceeded a certain threshold.
In addition to the non-linear compensation structure, our data offer another plank to identify how
financial incentives affect loan officers’ behavior. Specifically, the bank classified loans into six size and
business sector categories and calculated monthly bonuses by category. All loan officers in our final sample
handled loans in at least two of the loan categories used by the bank. Identification of the causal effect of
financial incentives on loan officer’s performance rests on comparing the behavior of the same loan officer
at the same point in time in a loan category were she was above the threshold with her behavior in another
loan category in which she was below the threshold. To achieve this specific comparison, we include loan
officer-by-month fixed effects in our most saturated regression specification.
We start with the observation that loan officers perform three different but interrelated tasks, all of
which affect the bank’s loan portfolio quality: they originate new loans (Heider and Inderst, 2012), screen
new loan applications, and monitor existing loans. Given the richness of our data, we are able to identify
1
See for example the Dodd–Frank Act, Title XIV, Subtitle A (loan origination of residential mortgages) and the Truth
in Lending Act (Regulation Z).

2
the causal effect of financial incentives on a loan officer’s effort related to these three tasks. We find that
loan officers that were at risk of losing their bonuses in the previous month focus on reducing the defaults
of the existing loans (“monitoring”) in the current month. This is rational in the sense that enhanced
monitoring should yield the highest “bang for the buck” by quickly improving portfolio performance.
The increased effort put into monitoring existing loans does not come at the expense of the effort spent
on originating loans and screening loan applications. Specifically, loan officers seem to increase their effort
in originating new loans, while their screening effort remains unchanged. This suggests that loan officers
see these three activities as complementary when trying to maximize their wage. The finding that screening
effort does not change is particularly interesting because it suggests that loan officers view screening as
relatively less important than the other two tasks when faced with a multitasking problem and a high-
powered incentive contract.
Our data and empirical setup do not allow us to comprehensively investigate whether the observed
behavior is optimal for the loan officers or for the bank. Two pieces of evidence we report suggest, however,
that at least the latter may not be the case. First, we show that loan officers, consistent with the incentives
from the contract they face, focus exclusively on the probability of default and disregard the loss given
default. In fact, the availabilityand the value of collateral do not enter the objective function of loan officers.
By focusing exclusively on the probability of default loan officers do not minimize the bank’s losses. This
is consistent with the idea that high-powered incentives to maximize short-term income present in
incomplete contracts carry the risk that loan officers neglect the tasks that are less well rewarded, but
nevertheless are in the interest of the bank (Holmström and Milgrom, 1991). Second, we do not find that
the observed loan behavior induced by the incentive contract leads to a reduction of the default risk of
newly originated loans, neither in the short-run nor in the medium-run. This suggests that the new loans
loan officers generate are not of higher quality.

3
Previous empirical work focuses primarily on the impact of performance-based compensation on loan
officers’ screening decisions.2
Most closely related to our paper, Cole et al. (2015) study three different
contract designs in a laboratory setting under real world conditions. They show that compensation that
rewards loan volume, but penalizes poor loan performance, entails more screening effort and a higher-
quality loan portfolio relative to other compensation schemes. Using data from a large U.S. commercial
lender, Agarwal and Ben-David (2013) study how loan-volume-based compensation affects the loan
volume and the delinquency rates. They find that when the compensation rewards volume, loan officers
generate more but lower-quality loans. Using data from a major European bank, Berg et al. (2013) study
how automated lending decisions based purely on hard information influence loan officers’ behavior when
the compensation depends on the loan volume generated and find that loan officers bias their assessment
of the borrowers’ risk to increase the pool of clients who are eligible to receive credit.
While this literature establishes an important causal relationship between financial incentives and
screening, it is largely mute about the multitasking problem that loan officers face in their job. This is
particularly problematic since based on our results as well as conversation with practitioners and academics
we consistently find that the loan officers’ multitasking problem includes three components: screening,
monitoring, and origination. Hence, studying these three components simultaneously is essential to get an
understanding of the loan officers’ response to incentive based contracts. Our setup is, to the best of our
knowledge, the only one suitable for this simultaneous analysis, making our paper an important contribution
to the existing literature. More specifically, our paper is the only one to emphasize the trade-off faced by
loan officers when they need to allocate effort into multiple tasks but are compensated according to an
incomplete contract in the sense that some activities are rewarded more than others.3
Our work also contributes to a strand of literature that analyzes other aspects of the role of loan officers
in financial institutions. For instance, Drexler and Schoar (2013) study the importance of relationships
2
Most of the literature on risk taking in banks focuses on top executives (e.g., Bebchuk and Spamann, 2010; Bolton
et al., 2010; Balachandran et al., 2011; Fahlenbrach and Stulz, 2011), rather than loan officers.
3
Our paper is also related to the literature on the presence of agency problems within banks (e.g., Liberti and Mian,
2009; Hertzberg et al., 2010).

4
between loan officers and borrowers for loan take-up and other loan outcomes. Fisman et al. (2015) use
data from India and show that cultural proximity matters for the efficiency of credit allocation. Qian et al.
(2015) use Chinese data to examine the effects of increased accountabilityof loan officers on the assessment
of credit risk. Finally, Beck et al. (2013) and Beck et al. (2015) analyze the impact of loan officer gender
on portfolio performance and gender-based discrimination.
2. Data and Identification Strategy
2.1 Institutional Background
Our data come from a large, for-profit, international commercial lender serving mainly individuals and
small- and medium-sized enterprises. The data set includes 37,533 loan applications and 27,742 loans
issued by the lender between January 1996 and October 2004. As the lender did not have any credit card
business in the sample period, all loans in our data set are either individual or small business loans. For all
loans, we observe the payment history and the point in time when a loan went into default. “Defaults” are
defined by the bank as a loan payment overdue for at least 30 days, consistent with the international
practice.4
This information enables us to construct a measure of monitoring because we have monthly
information on the default status of the loans.
The bank had 15 branches and 271 loan officers worked for it in the sample period. Loan officers have
full authority over their tasks: they independently process loans for their pre-existing clients as well as
actively seeking out new clients. In addition, they are responsible for monitoring the existing loans. For
example, if a loan payment is late, the loan officer can intensify the monitoring by calling the borrower,
sending her a letter, or visiting her to inquire about the reasons for the delay. To make monitoring salient,
loan officers can, for instance, threaten borrowers to deny them access to future credit or give them
unfavorable loan terms in subsequent loan applications.
4
Non-performing loans are taken out of a loan officer’s portfolio after being overdue for more than 60 days; these
loans are then taken care of by special workout units.

5
Loan applications by new borrowers are assigned to loan officers on a first-come, first-served basis.
New clients who walk into a branch are allocated to the loan officer who is available at the time and
assignment is not based on any particular characteristics of the loan officer or the borrower. The bank
categorizes loans into six loan categories that differentiate by loan size and the borrowers’ sector of business
activity. These six categories are small loans to private individuals for consumption purposes (up to 2,300
euros), very small business loans (up to 2,300 euros), small business loans (up to 10,000 euros), medium-
sized business loans (up to 50,000 euros), large business loans (bigger than 50,000 euros), and agricultural
loans. Loan officers can and do handle loans of more than one category at the same time.
The incentive-based compensation scheme was structured in such a waythat the bonus was proportional
to the loan officer’s lending volume as of the end of each month in a given loan category. However, in
months when the value-weighted defaults in the loan officer’s loan category-specific portfolio were above
3 percent, the bonus was cancelled for this loan category and month. Depending on the performance in each
loan category, the bonus was summed up over all well performing loan categories the loan officer was
covering and constituted up to a maximum of 150 percent of the loan officers’ fixed salary. Hence, the
incentive to keep the bonus was substantial under this performance-based contract.
2.2 Identification Strategy
Our identification strategy relies on two distinctive features of the compensation scheme used by the bank
in the sample period. The first feature is the non-linearity embedded in the incentive scheme, which foresees
that in months when the value-weighted default rate in a given loan category is above 3 percent, the bonus
is canceled. The second feature is that loan officers simultaneously handle loans in more than one category.
These two features allow us to compare the behavior of the same loan officer at the same point in time in a
loan category where she is at risk of losing her bonus (or already lost her bonus) to her behavior in another
category where she is not at risk of losing the bonus.
Specifically, we estimate variants of the following specification:
(1) yijt = 1Defaultsjct-1 + 2AtRiskjct-1 + 3AboveCutoffjct-1 + bX + A + eijt,

6
where yijt represents the different outcome variables that capture monitoring, loan origination, and screening
of loan i by loan officer j at month t; Defaultsjct-1 is the value-weighted average default rate of loans of
category c in the portfolio of loan officer j in month t-1 which controls for the linear effect of default on
each task’s effort; AtRiskjct-1 is a dummy variable that takes on the value of one if the value-weighted average
default rate of loans of category c in the portfolio of loan officer j in month t-1 was between 1.5 and 3
percent and captures the non-linear effect of being at risk of losing the bonus;5
and AboveCutoffjct-1 is a
dummy variable that takes on the value of one if the value-weighted average default rate of loans of category
c in the portfolio of loan officer j in month t-1 was above the cut-off of 3 percent. It captures the non-linear
effect of receiving no bonus because of a too high default rate.
In our specification the omitted reference benchmark comprises loan officers’ behavior in loan
categories with a value weighted average default rate below 1.5 percent, and therefore the coefficients of
interest 2 and 3 capture the effect of being at risk of losing the bonus and of being in the no bonus zone
respectively.
Depending on the estimation, X represents a vector of control variables at the loan (application)-level
(referred to as covariate set 1 in the regressions), loan officer characteristics that vary by time (covariate set
2a) and by loan category (covariate set 2b), and loan covariates that vary by time (covariate set 3). The
controls represented by X include the outstanding loan amount, the remaining time to maturity, the loan
officer experience measured as the total number of loans processed by loan officer j in loan category c since
she started working at the bank, and her workload measured as the number of outstanding loans in her
overall portfolio at time t. All these controls are described in detail in Table 1. Finally, A is a vector of fixed
effects. Standard errors in all regressions are clustered at the branch-month level since we expect the quality
of loans to be correlated with time and geographic location.6
5
We also performed all analyses using the intervals 1-3 percent and 2-3 percent as the at risk categories and results
are qualitatively very similar to the ones presented below.
6
Results are invariant to clustering at the loan officer level.

7
To validly compare the behavior of the same loan officer at the same point in time in loan categories
close or above the cut-off point with her behavior in a loan category where she was far away from the cut-
off point, in our most complete specification we include loan officer-by-month fixed effects, which control
for loan officer-specific time-varying unobserved factors that may influence their effort. In the monitoring
analyses, in our most saturated specification we also add loan fixed effects to control for any unobserved
variation in the portfolio composition.
For example, if in month t-1 a loan officer has a default rate of 1 percent in loan category A and of 2
percent in loan category B, the coefficient 2 would capture the difference between the monitoring effort
(or the other outcomes depending on the regression) of the loan officer on loans that are unlikely to put her
above the no-bonus threshold in time t and her monitoring effort on loans that are likely to put her above
the no-bonus threshold in time t. An important contribution in our approach is that it is immune to time-
varying loan officer characteristics which is likely an important source of bias in extant empirical work on
the subject. Furthermore, the non-linearity in the compensation scheme permit us to conduct a meaningful
placebo analysis using data from a period when the bank compensated its loan officers based on a fixed
wage contract.
Given our setup, we can only study loan officers who were active in more than one loan category during
the sample period. However, this is not a limitation since most loan officers in the sample satisfy this
condition. In particular, we only have to exclude 2,607 loan applications representing 7% of the sample.
2.3 Descriptive Statistics
All control variables used in the empirical analysis are defined in Table 1. In the regression tables, we
indicate which of the covariates are included in which specification. Table 2, Panel A provides descriptive
statistics for the four dependent variables used in the empirical analyses. Like much of the previous
empirical literature (see, for example, Mester et al., 2007, and Norden and Weber, 2010), we do not have
access to direct measures of loan officer monitoring, such as communication with the borrower, meetings,
phone calls, or emails. However, we can proxy monitoring effort based on status changes of the loans in a

8
loan officer’s loan portfolio. Our data set contains the one-to-one matching between borrowers and loan
officers, and for each loan we observe the application date, the issuing date and the date of any missed
payments. These data enable us to focus on the within-loan variation in monthly defaults. In line with the
literature, we assume that screening will only affect the overall and time-invariant riskiness of a loan, and
not its time series variation. Hence, we can attribute any effect of financial incentives that we find in this
set-up to changes in the extent (quality or intensity) to which loan officers monitor their borrowers.
Following this reasoning, we generate the variable ΔDefaultit which takes on the value of 1 if loan i was
not in default in the previous month t-1 but is in default in the current month t; it takes on the value of 0 if
there was no change in the default status from month t-1 to month t; and it takes on the value of -1 if loan i
was in default in the previous month but is not in default in the current month. This variable is constructed
on the loan-by-month level (it) so that the maximum number of observations per loan is equal to the time
to maturity of the loan measured in months minus one.
New application volumejct denotes the total new loan application volume in euro in loan category c
originated by loan officer j in month t. As in the case of monitoring, the true origination effort cannot be
observed and we use this variable as a proxy for the true origination effort. It is constructed on the loan
officer-by-loan category-by-month level (jct). Rejection ratep and Processing timep are our proxies for
screening effort. Rejection ratep is a dummy variable that takes on the value of one if loan application p
was rejected and Processing timep denotes the number of days the loan officer spent processing loan
application p. This variable is measured as the date of loan approval less the date of loan application. New
application volumejct and Processing timep are employed as 1 plus their natural logarithm in the regressions.
For the monitoring analyses, the sample consists of 244,294 observations at the loan-by-month level;
this is the number of approved loans (n = 27,742) summed over the number of months for which the loans
are present in the database during the sample period. The main outcome variable, ΔDefaultit, has a mean of
0.081 percent, which implies a slight deterioration of a loan’s credit quality over its lifetime. On average,
loan officers generated total loan amounts of 12,655 euros in a given loan category in a given month,
corresponding to 4 new loan applications and, hence, around 3,150 euros per loan application. The average

9
rejection rate was 26 percent and the average time to process a loan application was 6 days in the sample
period. As the latter three variables are constructed on different data levels, the sample sizes are
substantially smaller than the sample size used for the monitoring analyses.
PanelB of Table 2 shows descriptive statistics for the main explanatory variables used in the monitoring
analyses.7
On average, 0.4 percent of a loan officer’s loans in a given category defaulted in the previous
month, about 2.5 percent of the loan officers had a category-specific portfolio default rate of between 1.5
and 3 percent, and 3.5 percent had a category-specific portfolio default rate higher than 3 percent. On
average, loan officers had already processed 63 applications in a given loan category and the outstanding
loans, independent of the category, were 107. Finally, the average outstanding loan amount and loan
maturity on a monthly basis were 4,554 euros and 9 months, respectively.
3. Empirical Results
In this section, we show the main results on how financial incentives influence loan officers’ behavior in
the three tasks of their job: monitoring existing loans, origination of new loan applications, and screening
of new loan applications. We start by presenting univariate results for all dependent variables and then show
regression results for the four dependent variables in turn.
3.1 Univariate Results
Table 3 presents the unconditional difference estimates of the effect of the incentive-based compensation
plan for the loan officers’ three main tasks. Specifically, the first three columns show averages for the
month-to-month changes in defaults, the newly generated total loan application volume, the rejection rate
and the processing time in days conditional on the category-specific portfolio default frequency of the
respective loan officer in the previous months. We use the three intervals 0-1.5 percent for very well
7
We omit the corresponding descriptive statistics for the other analyses because results are qualitatively similar if we
collapse the loan-by-month level data to the other data levels used below (e.g., loan level). They are, however,
available upon request.

10
performing loan officers, 1.5-3 percent for loan officers that were at risk of losing their bonus, and above 3
percent for loan officers who did not receive a bonus in the respective loan category in the previous month.
The last two columns provide differences between well performing loan officers (column 1) and loan
officers who were at risk of losing their bonus (column 2) respectively loan officers who were above the
bonus threshold (column 3). Statistical inference in the last two columns of the table is based on running
OLS regressions without including controls, but clustering standard errors on the branch-month level.
We start by discussing the changes in monitoring, which are presented in the first row of Table 3. The
first column shows an average change in the default frequency of 0.09 percent for the loan officers with a
category-specific portfolio default frequency below 1.5 percent in the previous month. In contrast, the
second column indicates an average change in the default frequency of -0.17 percent for the loan officers
with a category-specific portfolio default frequency between 1.5 and 3 percent in the previous month. The
fourth column shows an unconditional difference of -0.26 percent between loan officers that were far away
from losing their bonus and loan officers that were at risk of losing their bonus (the difference between the
values given in columns 1 and 2). The difference is significant at the 1 percent level. The fifth column
displays an unconditional difference of -0.16 percent between loan officers that were far from losing their
bonus (column 1) and loan officers that lost their bonus (column 3) in the previous month. This difference
is also negative, but statistically insignificant. These results provide descriptive evidence that loan officers
seem to increase monitoring in the current month already when they were at risk of losing the bonus in the
previous month, but not significantly so anymore when they had already lost the bonus in the previous
month.
Loan officers seem to generate a larger new loan application volume and they seem to take longer to
process newly generated loan applications when they were in the at risk portfolio default region in month
t-1 relative to when they were far away from the no-bonus threshold. On the other hand, they do not seem
to reject significantly more loan applications. The results for the above cut-off region compared to the far
away from the cut-off region show that the difference is insignificant for the new loan application volume.

11
For the two screening proxies, we find that the rejection rate is significantly higher when loan officers are
above the no-bonus threshold in a given loan category and the processing time is also longer.
All in all, it appears that loan officers increase monitoring and engage more in loan origination, while
the univariate results do not draw a clear pattern as regards loan officers’ screening effort. While it seems
that the processing time increases, the results for the rejection rate are inconclusive. We explore and
comment further on this below.
3.2 Monitoring
We next estimate a number of variants of equation (1) with ΔDefaultit as the dependent variable to analyze
if the result from the univariate analysis holds if we include controls and fixed effects. The unit of
observation in these specifications is loan-by-month and the coefficients of interest are 2 and 3 from
equation (1). The former provides insight into whether loan officers react to the rising likelihood to lose the
bonus in subsequent months by increasing their monitoring effort, while the latter shows the effect on
monitoring effort after the loan officer has already lost her bonus in the previous month.
Column (1) of Table 4 presents the results without including any covariates or fixed effects, but
controlling for the linear effect of the loan officers’ loan category-specific portfolio default ratio of the
previous month. The AtRiskjct-1 variable is negative and significant and of a very similar magnitude as the
one from the univariate analysis. The AboveCutoffjct-1 variable continues to be negative, but insignificant.
These results do not change when we include different combinations of fixed effects and covariates. In the
most complete version of the model that includes loan and loan officer-by-month fixed effects to control
for borrower selection and time-varying unobservable loan officer characteristics (column (5)), we get a
significant coefficient of -0.0031 for the AtRiskjct-1 variable and an insignificant coefficient of the same size
for the AboveCutoffjct-1 variable. The magnitude in the former case amounts to an economically substantial
0.08 standard deviations of the monthly change in defaults.
These results suggest that financial incentives seem to induce loan officers to change their monitoring
behavior – they monitor more, more effectively, or both – in order to reduce the proportion of the portfolio

12
in default and maximize their wage. Moreover, loan officers do already react when they are somewhat close
to losing their bonus, while we find less compelling evidence for a change in monitoring for loan officers
that did not receive a bonus in the previous month.8
We next explore heterogeneous effects along three dimensions. First, we investigate whether loan
officers focus in their monitoring effort on loans which could be classified as risky loans according to their
estimated ex ante probability of default (PD). We develop a simple statistical model that predicts the ex
ante credit quality of selected loan applications using only the information available to the loan officer at
the time of the loan origination. The characteristics that are observable (or easily verifiable) for loan officers
(and for the econometrician) at the time of the loan origination are: the business sector of the borrower, the
applied loan amount, the leverage, the total assets, the cash over total assets, the applied loan amount over
total assets, whether the borrower has an account at the bank, whether the borrower has ever applied for a
loan at the bank before, the juridical form, and yearly country-specific macroeconomic variables like the
GDP, inflation, and unemployment rate. The macroeconomic variables used in this analysis are extracted
from the World Bank web page and are lagged by one year.
We then proceed to estimate an out-of-sample logit regression to explain the observed one-year-ahead
defaults for loans issued during the sample period. We recalibrate the model on a yearly basis to include
the most recent historical information. We use the coefficients obtained from these regressions to estimate
each borrower’s ex ante one-year-ahead PD.9
After estimating the expected PDs for all loans, we compute
the median PD per loan category over the entire sample period and create a dummy variable, Riskyi, that
takes on the value of one if the expected PD of a given loan is bigger than the median PD per loan category.
We then interact this dummy variable with the AtRiskjct-1 and the AboveCutoffjct-1 variables and estimate
equation (1) including these interactions. We only estimate the most saturated model corresponding to the
one from Table 4, column (5), that includes loan and loan officer-by-month fixed effects and further
8
We cannot rule out economically meaningful effects because of the large confidence intervals for the coefficient of
the AboveCutoffjct-1 variable.
9
These results are available on request.

13
covariates. Note that the time-invariant level term of the expected PD is subsumed in the loan fixed effects.
Table 5, column (1) shows the results.
The result for the AtRiskjct-1 and the AboveCutoffjct-1 variables are both not significant, but the interaction
between AtRiskjct-1 and the dummy variable indicating an expected PD bigger than the loan category-
specific median PD is negative, significant and quite large. This result suggests that loan officers increase
their monitoring effort particularly for those loans that could be classified as the ex ante riskier loans. One
explanation for this finding is that it may be relatively more effective to intensify monitoring for the ex ante
riskier loans because these loans have a higher propensity to go into default relative to the less risky loans.
Above, we show that loan officers change their monitoring behavior when they are at risk of losing
their bonus payments. This seems to be an intended consequence of the incentive contract and desirable
from the perspective of the bank. However, while the loan officers may be indifferent with regard to which
loans they monitor more when several are in default at the same time, from the perspective of the bank,
what matters is not the probability of default but rather the probability of default multiplied by the loss
given default. Does it matter that only the probability of default, not the loss given default, is incentivized
in the contract? While we do not have information on loan-by-loan loss given default, we do have some
indication of whether the loan was collateralized or guaranteed by a third party. We conjecture that on
average collateralized loans or loans that are guaranteed by a third party can be expected to have a lower
loss given default than other loans. Hence, we can test whether loan officers take collateralization into
account in their monitoring effort.
We construct a dummy variable, Unsecuredi, that takes on the value of one if a loan is secured neither
by a personal guarantee nor by mortgage collateral. We then estimate a variant of equation (1) in which we
include interaction terms with AtRiskjct-1 and AboveCutoffjct-1 with the newly created variable. The desired
monitoring behavior from the viewpoint of the bank would require the interaction terms to be negative and
significant, i.e. loan officers should focus on those loans that have the highest potential to generate large
losses for the bank.

14
Table 5, column (2) displays the results using the same specification as in the previous analysis of the
expected PD. None of the two interactions is significant, only the AtRiskjct-1 variable remains significant.
This finding suggests that there is no differential effect for secured versus unsecured loans or, in other
terms, loan officers do not take into account whether a loan is secured or not in their monitoring efforts.
Hence, while the observed loan officer behavior may be rational and optimal individually, this test suggests
that it may not necessarily be optimal for the bank. The evidence is consistent with the theoretical literature
on incentive-based contracts and multitasking. For example, Holmström and Milgrom (1991) show that
incentives in incomplete contracts carry the risk that agents will neglect those tasks that are less well
rewarded or are not part of the incentive structure at all, but nevertheless are in the interest of the bank.
Finally, we explore whether loan size influences monitoring effort. The metric that determines whether
or not the loan officer received a bonus is the value-weighted defaults in her portfolio. Hence, when faced
with losing the bonus, loan officers may rationally focus on large loans in their enhanced monitoring effort.
In addition, Cole et al. (2015) argue that the likelihood of improving the performance of a loan is more
limited in the case of small loans. Therefore, we expect the incentive effect on monitoring to be more
pronounced for large loans.
To test this effect, we compute a dummy variable, Largeit, that takes on the value of one if the
outstanding amount of loan i is above the average outstanding amount across all the loans a given category
in the sample period in month t. We then estimate an extended version of equation (1) that includes the
time-varying variable Largeit, as well as its interaction with the already-defined variables AtRiskjct-1 and
AboveCutoffjct-1.
Column (3) of Table 5 shows the results. The coefficient obtained on the interaction of
AtRiskjct-1*Largeit is positive but not significant, while the interaction with AboveCutoffjct-1 is negative and
significant. This suggests that the increase in the monitoring effort exerted by the loan officers above the
cut-off is stronger for larger loans, but loan size does not seem to matter when loan officers are at risk of
losing their bonus.

15
Taken together, these results support the idea that financial incentives affect loan officers’ monitoring
behavior. Loan officers exert more monitoring effort when the performance of their loan portfolio is
deteriorating enough so that they are concerned about losing the bonus payment. However, while the
observed behavior may be optimal individually, it may not necessarily be optimal for the bank because loan
officers do not seem to take aspects into account that should be important for the bank. We next test whether
the additional monitoring effort is due to an overall higher effort by loan officers when faced with losing
the bonus or whether it comes at the expense of the other two activities that loan officers typically perform:
origination and screening.
3.3 Loan Origination
We proxy loan origination effort by using the natural log of 1 plus the newly generated total loan application
volume of loan officer j in loan category c in month t. The data that we use for this test are at the loan
officer-loan category-month level and include 10,613 observations. As before, we estimate several variants
of equation (1) with this loan origination proxy as the dependent variable. The coefficients of interest are
again 2 and 3 in equation (1). We estimate equation (1) using OLS and cluster standard errors at the
branch-month level.
Table 6 presents the results. In column (1) that does not include any covariates or fixed effects, we find
a highly significant, positive coefficient for AtRiskjct-1 indicating that a loan officer who was at risk of losing
her bonus in the previous month in a given loan category generated significantly more new loan application
volume in this loan category relative to what she generated in another loan category in which she was far
away from losing her bonus in the previous month. The results do not change substantially, although the
coefficient size for AtRiskjct-1 is slightly reduced, when covariates and fixed effects are included in columns
(2) and (3). In column (4), we present results using the specification that includes loan officer-by-month
fixed effects and loan officer experience. As loan officers do not rotate across branches, this specification
also subsumes branch-by-month fixed effects that control for time-variant determinants of loan origination

16
at the branch level, such as regional changes in the loan demand.10
The results in this most saturated
specification do again not change and the economic magnitude of the effect amounts to an economically
significant 0.32 standard deviations of the new loan volume. On the other hand, the coefficient for the
AboveCutoffjct-1 dummy is never significant, with the exception of column (3) where we find a weakly
significant effect.
These findings indicate that an increase in monitoring effort does not come at the expense of reducing
loan origination effort, but that both activities are seen as complements from the loan officer’s perspective
in order to maximize the likelihood of receiving a bonus payment. Originating more new loan applications
makes sense from the perspective of the loan officer because a new loan is less likely to default at the
beginning of its maturity. Thus, a loan officer who lost her bonus or was at risk of losing her bonus is
lowering the likelihood of being above or close to the bonus threshold in the short run. We will investigate
in Section 3.6 whether this behavior increases ex post defaults over the loans’ maturity.
3.4 Screening
The third activity that we examine is screening effort. As in the case of monitoring, the true screening effort
is not observable, but we can proxy it by using the rejection rate and the time to process a loan application
in days. The latter variable is measured as the date of the loan approval decision less the date of the loan
application. We then estimate equation (1) with OLS using both variables as dependent variables. In these
tests, the data are at the loan application level, which allows us to estimate several variants of equation (1)
using different combinations of controls and fixed effects. The results are displayed in Table 7.
In Panel A of Table 7, we use the rejection rate as dependent variable. The table shows that neither the
coefficient for AtRiskjct-1 nor the coefficient for AboveCutoffjct-1 are significant in any of the specifications.
These results suggest that while loan officers seem to increase effort to originate more loan volume, they
do not change their behavior with regard to the decision to accept or reject a loan application.
10
Note that in this specification it is not possible to include loan fixed effects because we do not explore the time-
series variation within each loan.

17
On the other hand, in Panel B of Table 7, we find that the processing time increases as evidenced by
the significant coefficient of the AtRiskjct-1 variable in all four specifications of the table. The coefficient of
AboveCutoffjct-1 is only significant in the first column where we do not include any controls or fixed effects
apart from the linear effect of a loan officer’s default ratio in a given loan category in the previous month.
However, once we add controls and fixed effects, this effect goes away, while the coefficient of AtRiskjct-1
remains positive and significant, even though in the most complete specification in column (4), statistical
significance is somewhat reduced and the coefficient is much smaller than in column (1) where no controls
and fixed effects are included. The economic effect in this last specification amounts to 0.10 standard
deviations of the processing time.
The results reported so far suggest that loan officers increase effort for the monitoring and origination
task in order to maximize the likelihood of receiving a bonus payment. This is novel evidence that goes
beyond the findings in Cole et al. (2015) and Agarwal and Ben-David (2013). The effect of financial
incentives on loan officers’ screening effort are less clear in our setting. On the one hand, we find that loan
officers do not reject or accept more loan applications in an attempt to recover their bonus payment. On the
other hand, the processing time increases. However, the latter finding could simply be a mechanical
consequence of the bigger workload stemming from originating more loan application volume and
increasing monitoring effort, which could result in each loan application taking more time to be processed.
Interestingly, our evidence clearly suggests that loan officers already react to the financial incentives
when they are at risk of losing their bonus, but when they lost their bonus in the previous month, they do
not seem to change effort significantly for any of the three tasks explored in this study. To the extent that
this was the intended purpose of the given incentive contract and desired by the bank, this would seem to
be optimal behavior from the viewpoint of the bank. However, we also document one effect that may not
be desirable from the viewpoint of the bank: loan officers do not focus on those loans that are not
collateralized. This should clearly not be in the interest of the bank. Such a result also shows the
incompleteness of incentive contracts when it is not possible to contract on all relevant dimensions that
would make loan officer behavior optimal for the bank.

18
Our results do not inform whether loan officers solve their time allocation problem rationally and
optimally from their individual point of view. We document that effort increases along two of their tasks,
which suggests that loan officers see their activities as complements rather than substitutes. However,
focusing on one activity exclusively at the expense of the other two activities may increase the likelihood
of receiving a bonus payment more relative to increasing effort along any of the other tasks. Our empirical
setup does not allow us to analyze this type of question and further research is needed to understand this
better.
3.5 Placebo Test
To address any remaining identification concerns, this section discusses the results of a placebo test that
uses data from a period in which the bank no longer made use of an incentive-based payment contract. In
such an environment, we would also not expect to find any results if our findings reported above were really
driven by financial incentives, unless financial incentives were substituted perfectly by some other
incentives (for instance, changes to when loan officers are fired or promoted) that implicitly or explicitly
use the same cut-off threshold. We view this as highly unlikely, but cannot fully rule out such a possibility.
Also, private conversations with bank management did not point towards such a “substitution effect”.
For this placebo test, we make use of a two-step change in the compensation scheme implemented by
the bank from November 2004 onwards. Between November 2004 and December 2005, the bonus payment
was limited to a maximum of 50 percent of the fixed salary. From January 2006 onwards, the incentive
compensation was completely abolished and substituted by a fixed salary only regime. We collect data from
the time period January 2006 until September 2007 and run all regressions for the four dependent variables
from Tables 4, 6, and 7 using the most saturated specification in each case.11
Rather than the data being
from a different time period, everything else is identical in these regressions. Table 8 shows the results. In
11
We do not include the observations from the transition period from November 2004 to December 2005 in which the
bonus was limited to 50 percent of the fixed salary because a priori it is not clear what effects to expect in that time
period.

19
line with our expectation, none of the estimates in any of the four columns is statistically different from
zero. This suggests that the results reported above are indeed capturing the causal effect of financial
incentives on loan officer behavior.
3.6 Loan Performance
Loan officers intensify monitoring, they originate more loans, but do not seem to change their screening
effort substantially when they are at risk of losing their bonus. In order to understand better whether this
ultimately results in riskier or less risky loan portfolios, we evaluate the ex post quality of the loan portfolio
of loan officers using different evaluative windows. This test uses all approved loans and their observed
defaults during four evaluative windows: in the first four months after the loan was granted, between the
5th
and the 8th
month, between the 9th
and 12th
month, and after more than 12 months as the dependent
variable in estimating equation (1). We use the same specification as the one in Table 7, column (4) that
includes loan category fixed effects, loan officer-by-month fixed effects, and the appropriate covariates.
The results of this test are shown in Table 9.
The table shows that AtRiskjct-1 and AboveCutoffjct-1 are not significant in any of the evaluative windows.
Most of the coefficients are indeed positive so if anything, the observed loan officer behavior tends to
increase the default likelihood. While the increase in default likelihood is not significant in statistical terms,
this finding may indicate that the loans which the loan officers generate when they are at risk of losing their
bonus, may not necessarily be loans of good quality.12
Hence, while the change in behavior may maximize
loan officers’ wage payments in the short-run, it may not be optimal from the bank’s perspective.
4. Conclusion
We study how the behavior of loan officers at a large international bank varies depending on a performance-
based compensation plan. What makes our setting particularly interesting is that the compensation plan
12
This is consistent with the finding that the rejection rate does not change even though loan officers generate more
loan application volume when they are at risk of losing their bonus.

20
studied is highly non-linear. In particular, it rewards the loan officers with a bonus that increases
monotonically with the loan volume as long as the proportion of the portfolio in default in a given loan
category is below 3 percent, but the bonus is canceled once the proportion of the portfolio in default
surpasses that threshold. Furthermore, our empirical setup and data allow us to compare the behavior of the
same loan officer at the same point in time when she is at risk of losing her bonus in one loan category as
opposed to another loan category in which she is far away from the risk of losing her bonus.
The results indicate that when loan officers were at risk of losing their bonus in the previous month,
they allocate more time to two of the three activities of a loan officer in the current month: monitoring and
origination. Interestingly, screening effort as proxied by the rate of rejection of posted loan applications
does not seem to change. This suggests that loan officers see the other two activities as relatively more
important (or easier to change in the short-run) than screening in order to maximize their wages.
On the other hand, having lost the bonus in the previous month does not lead to an increase in effort for
any of these three tasks. The finding that loan officers do not behave in a myopic way may be interpreted
as rational behavior although our empirical setup does not allow us to analyze whether the documented
behavior is optimal individually. For instance, loan officers could be better off if they focused exclusively
on monitoring at the expense of the other two activities because monitoring should, at least in the short-run,
have the biggest relative effect on their portfolio default risk.
On the other hand, the evidence also shows that such incentive-based contracts, by rewarding certain
activities over others, may not necessarily result in outcomes that are preferable from a bank perspective.
Specifically, we find that loan officers do not concentrate their monitoring activities on loans that are less
collateralized, although this should be clearly in the interest of the bank. One bank-managerial implication
of our results is to contract on all dimensions that are relevant for the bank, but any incentive-based contract
is necessarily incomplete and therefore may reward some activities “too little” that would also be in the
interest of the bank. For example, loan officers rationally focus on the probabilities of default, rather than
the loss to the bank, which is a combination of default probability and loss given default. One theoretically

21
optimal solution to this problem could be to give loan officers an equity stake in the loan as in Cole et al.
(2015), but this is hardly feasible in practice.
While our empirical setup bars us from making statements about the optimality of the observed loan
officer behavior for the bank, the evidence shows that the increased loan officer effort with regard to
origination of new loan application volume does not improve portfolio performance in the short- and the
medium-run. This may be another indication that an incentive contract like the one studied in this paper,
does not necessarily imply optimal behavior from the bank’s perspective.

22
References
Agarwal, S., Ben-David, I., 2013, Do Loan Officers’ Incentives Lead to Lax Lending Standards?
Unpublished working paper.
Balachandran, S., Kogut, B., Harnal, H., 2011, Did Executive Compensation Encourage Extreme Risk-
Taking in Financial Institutions? Unpublished working paper.
Bebchuk, L., Spamann, H., 2010, Regulating Bankers’ Pay, Georgetown Law Journal 98, 247–287.
Beck, T., Behr, P., Guettler, A., 2013, Gender and Banking: Are Women Better Loan Officers? Review
of Finance 17, 1279–1321.
Beck, T., Behr, P., Madestam, A., 2015, Sex and Credit: Is There a Gender Bias in Lending?
Berg, T., Puri, M., Rocholl, J., 2013, Loan Officer Incentives and the Limits of Hard Information,
Bolton, P., Mehran, H., Shapiro, J., 2010, Executive Compensation and Risk-Taking, Unpublished
working paper.
Cole, S., Kanz, M., Klapper, L., 2015, Incentivizing Calculated Risk-Taking: Evidence from an
Experiment with Commercial Bank Loan Officers, Journal of Finance 70, 537-575.
Drexler, A., Schoar, A., 2013, Do Relationships Matter? Evidence from Loan Officer Turnover,
Management Science 60, 2722-2736.
Fahlenbrach, R., Stulz, R., 2011, Bank CEO Incentives and the Credit Crisis, Journal of Financial
Economics 99, 11–26.
Fisman, R., Paravisini, D., Vig, V., 2015, Cultural Proximity and Loan Outcomes, Unpublished
working paper.
Heider, F., Inderst, R., 2012, Loan Prospecting, Review of Financial Studies 25, 2381–2415.
Hertzberg, A., Liberti, J.M., Paravisini, D., 2010, Information and Incentives inside the Firm:
Evidence from Loan Officer Rotation, Journal of Finance 65, 795–828.

23
Holmström, B., Milgrom, P., 1991, Multi-task Principal-agent Analysis: Incentive Contracts, Asset
Ownership, and Job Design, Journal of Law, Economics, and Organization 7, 24–52.
Liberti, J.M., Mian, A., 2009, The Effect of Hierarchies on Information Use, Review of Financial
Studies 22, 4057–4090.
Mester, L., Nakamura, L., Renault, M., 2007, Transactions Accounts and Loan Monitoring, Review
of Financial Studies 20, 529–556.
Norden, L., Weber, M., 2010, Credit Line Usage, Checking Account Activity, and Default Risk of
Bank Borrowers, Review of Financial Studies 23, 3665–3699.
Qian, J., Strahan, P.E., Yang, Z., 2015, The Impact of Incentives and Communication Costs on
Information Production: Evidence from Bank Lending, Journal of Finance 70, 1457–1493.

24
Table 1: List of covariates
This table shows the covariates that are used in the different regressions.
Covariate Description
Covariate set 1: Loan (application) level
Leverage Total liabilities/total assets
Cash over total assets Liquid assets/total assets
Total assets In euros
ln(Applied amount) Natural logarithm of the loan size applied for by the borrower in euros
ln(Applied maturity) Natural logarithm of the loan maturity the borrower applied for in days
Applied loan over total assets Loan size applied for by the borrower in euros/total assets
Juridical form business 1 if the client is a legal entity and 0 if the client is a natural person
Available account
1 if the client has other accounts (checking, savings, etc.) at the bank at the time
of the loan application and 0 otherwise
Guarantee 1 if the client provides personal or mortgage guarantees and 0 otherwise
Has been in default 1 if the client has been in default with a previous loan
Has been rejected 1 if the client had submitted a previous loan application that was rejected
Last week of the month 1 for loans applied for in the last week of the month and 0 otherwise
Number of loan applications 1 for the first loan application, 2 for the second loan application, etc.
Past experience 1 if the borrower had past experience with the loan officer and 0 otherwise
Loan category Size- and sector-specific categories
Loan destination
Loan used for working capital, fixed assets, mixed working capital and fixed
assets, real estate, consuming, or others
Business sector Agriculture, production, construction, transport, trade, other services, or others
Covariate set 2a: Loan officer*time level
Number of outstanding loans Number of outstanding loans per loan officer (approximates workload)
Covariate set 2b: Loan officer*loan category*time level
Loan officer experience
Number of loan applications that were already handled by a loan officer in a
given loan category
Covariate set 3: Loan*month level
ln(Outstanding amount) Natural logarithm of the outstanding loan amount in euros
ln(Remaining maturity) Natural logarithm of the remaining maturity in months

25
Table 2: Descriptive statistics
This table shows the descriptive statistics for the main variables of interest. Panel A shows descriptive statistics for the main dependent variables. ΔDefaultit takes
the value of 1 if loan i was not in default in the previous month but is in default in the current month; it takes the value of 0 if there was no change in the default
status; and it takes the value of -1 if loan i was in default in the previous month but is not in default in the current month. This variable for the monitoring analysis
is on the loan-by-month level. New application volumejct denotes the total new loan application volume in euros in loan category c originated by loan officer j in
month t. This variable for the origination analysis is on the loan officer-by-loan category-by-month level (jct). Rejection ratep is a dummy variable that equals 1 if
loan application p was rejected. Processing timep denotes the number of days the loan officer spent processing loan application p. New application volumejct and
Processing timep are employed by 1 plus their natural logarithm in the regressions. Panel B comprises the main explanatory variables for the monitoring analysis.
Defaultsjct-1 are the loan officer j’s average loan portfolio defaults in loan category c in month t-1. AtRisk jct-1 is a dummy variable that takes the value of one if the
defaults in loan officer j’s portfolio in loan category c were between 1.5 and 3 percent in month t-1. AboveCutoffjct-1 is a dummy variable that takes the value of one
if the defaults in loan officer j’s portfolio in loan category c were above the cut-off value of 3 percent in month t-1. Loan officer experiencejct is the number of loan
applications in loan category c handled by loan officer j until month t. Number of outstanding loansjt is loan officer j’s number of outstanding loans in month t,
which we use as a proxy for the loan officer’s workload. The last two rows present time-variant loan characteristics. Outstanding amountit is loan i’s outstanding
amount in euros in month t. Remaining maturityit is the remaining maturity of loan i in month t, which is denoted in months. The latter two variables are used by 1
plus their natural logarithm in the regressions.
Variable Data level Mean N Std. dev. Min. p50 Max.
Panel A: Main dependent variables
ΔDefaultit Loan*month 0.00081 244,294 0.03948 -1.00000 0.00000 1.00000
New application volumejct Officer*category*month 12,655 10,613 24,468 0 4,563 461,742
Rejection ratep Application 0.261 37,533 0.439 0.000 0.000 1.000
Processing timep Application 6 37,533 6 0 4 20
Panel B: Main explanatory variables for the monitoring analysis
Defaultsjct-1 Officer*category*month 0.00401 244,294 0.0240 0.0000 0.0000 1.0000
AtRiskjct-1 Officer*category*month 0.02454 244,294 0.1547 0.0000 0.0000 1.0000
AboveCutoffjct-1 Officer*category*month 0.03495 244,294 0.1837 0.0000 0.0000 1.0000
Loan officer experiencejct Officer*category*month 63 244,294 51 1 48 281
Number of outstanding loansjt Officer*month 107 244,294 66 1 99 387
Outstanding amountit Loan*month 4,554 244,294 11,070 0 1,656 340,866
Remaining maturityit Loan*month 9 244,294 9 0 7 88

26
Table 3: Univariate results
This table shows univariate results for all four outcome variables. Variables are defined in Table 2. The first column
provides the averages for loan officers far below the bonus threshold, i.e., for a default frequency in a given loan
category in the previous month below 1.5 percent. The second column shows the averages for loan officers close to
the bonus threshold, i.e., if a loan officer’s default frequency in a given loan category was between 1.5 and 3 percent
in the previous month. The third column provides the averages for loan officers above the bonus threshold, i.e., if a
loan officer’s default frequency in a certain loan category was above 3 percent in the previous month. Statistical
inference for the test of differences in the last two columns is based on OLS regressions that use standard errors
clustered at the branch-month level. *, **, and *** denote significance at the 10, 5, and 1 percent level, respectively.
AtRisk AboveCutoff Differences
Default frequency: I: 0-1.5% II: 1.5%-3% III: > 3% II-I III-I
Monitoring
ΔDefaultit 0.0009 -0.0017 -0.0007 -0.0026*** -0.0016
Origination
In(New application volumejct) 7.43 8.94 7.56 1.51*** 0.13
Screening
Rejection ratep 0.2596 0.2256 0.3337 -0.0340 0.0741***
ln(Processing timep) 1.59 1.96 1.96 0.37*** 0.37***

27
Table 4: Monitoring – Baseline results
This table shows the results of the OLS regression
ΔDefaultit = 1Defaultsjct-1 + 2AtRiskjct-1 + 3AboveCutoffjct-1 + bX + A + eit,
and examines whether an incentive-based compensation plan affects the changes in monitoring effort. Variables are
defined in Table 2. The covariate sets (X) are defined in Table 1; the first set includes time-invariant covariates at the
loan-level, the second set includes time-variant loan officer characteristics, and the third set includes time-variant
covariates at the loan-level. We gradually add fixed effects on the loan category (c), the time (t), the loan officer (j),
the loan (i) and the loan officer-by-month (jt) levels in columns (2) to (5). *, **, and *** denote significance at the
10, 5, and 1 percent level, respectively. Standard errors in parentheses are clustered at the branch-month level.
(1) (2) (3) (4) (5)
Defaultsjct-1 -0.0162 -0.0793*** -0.1102*** -0.1501*** -0.1807***
(0.0225) (0.0216) (0.0203) (0.0298) (0.0371)
AtRiskjct-1 -0.0023** -0.0033*** -0.0039*** -0.0042*** -0.0031**
(0.0009) (0.0010) (0.0010) (0.0012) (0.0015)
AboveCutoffjct-1 -0.0002 -0.0019 -0.0024 -0.0035 -0.0031
(0.0018) (0.0019) (0.0019) (0.0027) (0.0026)
Loan category fixed effects No Yes Yes No No
Month fixed effects No No Yes Yes No
Loan officer fixed effects No No Yes No No
Loan fixed effects No No No Yes Yes
Loan officer*Month fixed effects No No No No Yes
Covariate set 1 No Yes Yes No No
Covariate set 2 No Yes Yes Yes Yes (2b)
Covariate set 3 No Yes Yes Yes Yes
Data level Loan*month Loan*month Loan*month Loan*month Loan*month
Observations 244,294 244,294 244,294 244,294 244,294
Adj. R square 0.0002 0.0411 0.0477 0.0028 0.0205

28
Table 5: Monitoring – Heterogeneous effects
This table shows the results of the OLS regressions
ΔDefaultit = 1Defaultsjct-1 + 2AtRiskjct-1 + 3AboveCutoffjct-1
+ 4AtRiskjct-1*LoanTypei + 5AboveCutoffjct-1*LoanTypei + bX + ai + ajt + eit
ΔDefaultit = 1Defaultsjct-1 + 2AtRiskjct-1 + 3AboveCutoffjct-1 + 4AtRiskjct-1*LoanTypeit
+ 5AboveCutoffjct-1*LoanTypeit + 6LoanTypeit + bX + ai + ajt + eit,
and examines whether the increase in the monitoring effort of loan officers who came close to or surpassed the cut-
off was focused on specific types of loans. We use the most saturated specification of Table 4 with loan (i) and loan
officer-by-month (jt) fixed effects. The variables are defined in Table 2, except for the interaction terms between
AtRisk and AboveCutoff and the respective loan type. In the first column, we define a loan as ex ante risky if the
estimated default risk was above the median in loan category c. We measure a loan’s default risk as the predicted ex
ante credit risk based on historical information. The credit risk measure is annually calibrated using variables that are
observable (or easily verifiable) for loan officers at the time of loan origination: the business sector of the borrower,
the loan amount needed, the leverage, the total assets, the cash over total assets, the applied loan amount over total
assets, whether the client had an account at the bank, whether the client had ever applied for a loan at the bank, the
juridical form, and the three yearly macroeconomic variables GDP, inflation, and unemployment. The ex ante
probability of default is based on a logit regression using the aforementioned variables. This variable does not vary
over time and its level term is thus saturated by the loan fixed effect. In the second column, we use the dummy variable
Unsecured, which is defined as 1 – Guarantee and takes the value of one if a loan does not come with a personal
and/or mortgage guarantee. This variable does not vary over time and its level term is thus saturated by the loan fixed
effect. In the third column, the dummy variable Large takes the value one if a loan’s outstanding amount is above the
median outstanding loan amount of the current month. This variable does vary over time and its level term is thus
included in the estimation. The covariate sets are defined in Table 1. *, **, and *** indicate significance at the 10, 5,
and 1 percent level, respectively. Standard errors in parentheses are clustered at the branch-month level.
(1) (2) (3)
Interaction variable (loan type): Risky Unsecured Large
Defaultsjct-1 -0.1802*** -0.1793*** -0.1806***
(0.0371) (0.0372) (0.0370)
AtRiskjct-1 -0.0003 -0.0037** -0.0037*
(0.0016) (0.0019) (0.0019)
AboveCutoffjct-1 -0.0013 -0.0014 0.0014
(0.0034) (0.0033) (0.0032)
AtRiskjct-1*LoanTypei(t) -0.0050** 0.0008 0.0013
(0.0023) (0.0023) (0.0021)
AboveCutoffjct-1*LoanTypei(t) -0.0031 -0.0025 -0.0106***
(0.0034) (0.0038) (0.0033)
LoanTypeit 0.0002
(0.0003)
Loan fixed effects Yes Yes Yes
Loan officer*month fixed effects Yes Yes Yes
Covariate set 2b Yes Yes Yes
Covariate set 3 Yes Yes Yes
Data level Loan*month Loan*month Loan*month
Observations 244,294 244,294 244,294
Adj. R square 0.0206 0.0205 0.0209

29
Table 6: Loan origination
In(New application volumejct) = 1Defaultsjct-1 + 2AtRiskjct-1 + 3AboveCutoffjct-1 + bX + A + ejct,
and examines whether an incentive-based compensation plan affects the loan origination effort. The data are at the
loan officer-by-loan category-by-month level (jct). We measure loan origination by In(New application volumejct)
which denotes 1 plus the natural logarithm of the total new loan application volume in loan category c originated by
loan officer j in month t. All other variables are defined in Table 2. The regressions control for the indicated set of
covariates (see Table 1) and include the indicated sets of fixed effects. *, **, and *** indicate significance at the 10,
5, and 1 percent level, respectively. Standard errors in parentheses are clustered at the branch-month level.
(1) (2) (3) (4)
Defaultsjct-1 -0.4648 0.1362 1.3701 2.3016
(1.8928) (1.9288) (1.2091) (1.7299)
AtRiskjct-1 1.5174*** 1.0644*** 1.0260*** 1.0810***
(0.2388) (0.2369) (0.2310) (0.3267)
AboveCutoffjct-1 0.1822 0.0094 0.4239* 0.0447
(0.2936) (0.2945) (0.2434) (0.3255)
Fixed effects
Month No No Yes No
Loan officer No No Yes No
Loan officer*Month No No No Yes
Covariate set 2b No Yes Yes Yes
Data level Officer*category
*month
Officer*category
*month
Officer*category
*month
Officer*category
*month
Observations 10,613 10,613 10,613 10,613
Adj. R square 0.0024 0.0208 0.2483 0.3219

30
Table 7: Screening
Yp = 1Defaultsjct-1 + 2AtRiskjct-1 + 3AboveCutoffjct-1 + bX + A + ep,
and examines whether an incentive-based compensation plan affects the rejection rate and the loan application
processing time. The data are at the loan application level (p). In Panel A, the dependent variable, Rejection ratep, is
a dummy variable that equals 1 if loan application p was rejected. In Panel B, the dependent variable, Processing
timep, is the processing time of loan application p. It is measured by1 plus the natural logarithm of the number of days
a loan officer spent evaluating a loan application. All other variables are defined in Table 2. The regressions control
for the indicated sets of covariates (see Table 1) and include the indicated sets of fixed effects. *, **, and *** indicate
significance at the 10, 5, and 1 percent level, respectively. Standard errors in parentheses are clustered at the branch-
month level.
Panel A: Rejection ratep
(1) (2) (3) (4)
Defaultsjct-1 0.2646 -0.0602 -0.0885 -0.0932
(0.2013) (0.0818) (0.0594) (0.1007)
AtRiskjct-1 -0.0396 -0.0133 -0.0014 0.0006
(0.0248) (0.0095) (0.0090) (0.0123)
AboveCutoffjct-1 0.0469 0.0121 0.0000 0.0062
(0.0304) (0.0122) (0.0102) (0.0133)
Loan category fixed effects No Yes Yes Yes
Month fixed effects No No Yes No
Loan officer fixed effects No No Yes No
Loan officer*Month fixed effects No No No Yes
Covariate set 1 No Yes Yes Yes
Covariate set 2 No Yes Yes Yes (2b)
Data level Application Application Application Application
Observations 37,533 37,533 37,533 37,533
Adj. R square 0.0009 0.7615 0.7757 0.7839
Panel B: ln(Processing timep)
(1) (2) (3) (4)
Defaultsjct-1 1.5939*** 0.4127 -0.0700 -0.0615
(0.4874) (0.2907) (0.1589) (0.3252)
AtRiskjct-1 0.3385*** 0.1907*** 0.0973*** 0.0951**
(0.0597) (0.0462) (0.0353) (0.0465)
AboveCutoffjct-1 0.2092** 0.0656 -0.0099 0.0271
(0.0810) (0.0530) (0.0358) (0.0478)
Loan category fixed effects No Yes Yes Yes
Month fixed effects No No Yes No
Loan officer fixed effects No No Yes No
Loan officer*Month fixed effects No No No Yes
Covariate set 1 No Yes Yes Yes
Covariate set 2 No Yes Yes Yes (2b)
Data level Application Application Application Application
Observations 37,533 37,533 37,533 37,533
Adj. R square 0.0075 0.2976 0.4871 0.5336

31
Table 8: Placebo test
This table shows the OLS regressions of a placebo test in which we repeat the estimations from the Tables 4, 6, and 7 considering observations from the no-bonus
period ranging from January 2006 to September 2007. We use the most saturated specifications of the Tables 4, 6, and 7. Covariates and fixed effects are included
in the regressions as specified in the table. All variables are defined in Table 2.*, **, and *** indicate significance at the 10, 5, and 1 percent level, respectively.
Standard errors in parentheses are clustered at the branch-month level.
Monitoring Origination Screening
ΔDefaultit In(New application volumejct) Rejection ratep ln(Processing timep)
Defaultsjct-1 -0.1799*** -0.8747 -0.1234 -0.0882
(0.0377) (2.7786) (0.1157) (0.2975)
AtRiskjct-1 -0.0015 0.4459 0.0067 -0.0159
(0.0010) (0.3131) (0.0101) (0.0276)
AboveCutoffjct-1 -0.0007 0.6152 0.0108 0.0259
(0.0024) (0.4327) (0.0149) (0.0402)
Loan category fixed effects No No Yes Yes
Loan fixed effects Yes No No No
Loan officer*Month fixed effects Yes Yes Yes Yes
Covariate set 1 No No Yes Yes
Covariate set 2b Yes Yes Yes Yes
Covariate set 3 Yes No No No
Data level Loan*month Officer*category*month Application Application
Observations 462,622 8,278 41,312 41,312
Adj. R square -0.0159 0.3954 0.7451 0.3555

32
Table 9: Ex post loan performance
DefaultLikelihoodi = 1Defaultsjct-1 + 2AtRiskjct-1 + 3AboveCutoffjct-1 + bX + ac + ajt + ei,
and examines whether an incentive-based compensation plan affects the ex post loan performance. The data are at the
loan-level (i). The default likelihood is 1 if a loan missed a payment for more than 30 days at least once within various
time periods after the loan was granted. In the first column we concentrate on the first four months after the loan was
granted, in the second column we use month 5 to 8, in the third column months 9 to 12, and in the last column all
observations with a maturity over 12 months. The regressions control for the covariate sets 1 and 2b (see Table 1) as
well as loan category (c) and loan officer-by-month (jt) fixed effects. All other variables are defined in Table 2. *, **,
and *** indicate significance at the 10, 5, and 1 percent level, respectively. Standard errors in parentheses are clustered
at the branch-month level.
Maturity range (in months) 1-4 5-8 9-12 > 12
Defaultsjct-1 -0.0100 0.5498*** 0.4534*** 0.3552
(0.0303) (0.1616) (0.1694) (0.2687)
AtRiskjct-1 0.0071 0.0008 0.0110 0.0190
(0.0054) (0.0128) (0.0097) (0.0230)
AboveCutoffjct-1 0.0023 -0.0043 0.0204 -0.0050
(0.0059) (0.0136) (0.0142) (0.0217)
Loan category fixed effects Yes Yes Yes Yes
Loan officer*month fixed effects Yes Yes Yes Yes
Covariate set 1 Yes Yes Yes Yes
Covariate set 2b Yes Yes Yes Yes
Data level Loan Loan Loan Loan
Observations 30,388 20,680 13,593 6,950
Adj. R square 0.1533 0.3225 0.4375 0.3017

Financial incentives and loan officer behavior: multitasking and allocation of effort under an incomplete contract

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Similar to Financial incentives and loan officer behavior: multitasking and allocation of effort under an incomplete contract

Similar to Financial incentives and loan officer behavior: multitasking and allocation of effort under an incomplete contract (20)

More from FGV Brazil

More from FGV Brazil (20)

Recently uploaded

Recently uploaded (20)

Financial incentives and loan officer behavior: multitasking and allocation of effort under an incomplete contract