AGRICULTURAL COMMODITY PRICES AND
LOAN LOSSES AT A FARM CREDIT
ASSOCIATION: AN AUTOREGRESSIVE
DISTRIBUTED LAG ANALYSIS
A Thesis
Presented to the Faculty of the Graduate School
of Cornell University
in Partial Fulfillment of the Requirements for the Degree of
M.S.
by
Fenni Lian
May 2020
©c 2020 Fenni Lian
ALL RIGHTS RESERVED
ABSTRACT
This study investigates how agricultural commodity prices affect loan loss pro-
visions (LLP) and loan loss charge-offs (LLC) in a financial institution of the
Farm Credit System. Commodity prices have an impact on farmers’ revenue
generation, which might influence farmers’ ability to repay loans. This study
hypothesizes that loan loss provisions are a function of loan loss charge-offs
and commodity prices, and alternatively, that loss provisions and commodity
prices determine charge-offs. An autoregressive distributed lagged model is es-
timated for a financial institution of the Farm Credit System using data from
2010 to 2019. The empirical evidence suggests that LLP is positively related to
LLC and various commodity prices, with lagged components of different peri-
ods. LLC is also positively correlated with LLP and various commodity prices,
with lagged components of different periods.
Keywords: agriculture, commodity, loan loss, Farm Credit, time series anal-
ysis
BIOGRAPHICAL SKETCH
Fenni received her bachelor’s degree in Beijing, China, in the accounting field.
The passion for exploring underlying economic theories inspired her to con-
tinue studying in the master of science program in Applied Economics at Cor-
nell University. Her keen interest in statistical modeling motivated her the rela-
tionship between commodity prices and loan losses in the farm credit system in
the States. To finish this thesis, she also seeked help from the statistics depart-
ment at Cornell. Fenni will be working as a data scientist and further pursue
her interest in statistics and economics in the industry upon graduation.
iii
This document is dedicated to all Cornell graduate students.
iv
ACKNOWLEDGEMENTS
I appreciate the weekly meetings with professor Loren Tauer from Dyson
School, Cornell University. I learned not only economic knowledge but also
economic intuition from the professor. This thesis cannot be completed without
the guidance from professor Tauer.
I am also grateful to professor Sumanta Basu from the Statistics Department,
Cornell University. He encouraged me to develop technical time series model-
ing skills.
Last but not least, I hold the most sincere gratefulness towards my parents,
who supported and sponsored me to pursue my studies at Cornell University.
v
TABLE OF CONTENTS
Biographical Sketch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
1 Introduction 1
2 Literature Review 5
3 Method 9
3.1 Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.1.1 Charge-offs, Provisions and Commodities . . . . . . . . . . 9
3.1.2 Autoregressive . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.1.3 Lagged Effect . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.2 ARDL Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3 Sample and Data Description . . . . . . . . . . . . . . . . . . . . . 12
3.3.1 Data Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.3.2 Data Processing . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.4 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4 Results and Forecasting 25
4.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.1.1 Charge-offs . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.1.2 Provisions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.2 Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2.1 Charge-offs . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2.2 Provisions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.2.3 Comparison to the Actual Number . . . . . . . . . . . . . . 37
5 Conclusion and Suggestions 38
5.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.2 Suggestions for Future Work . . . . . . . . . . . . . . . . . . . . . . 39
5.2.1 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.2.2 More Cases from Farm Credit System . . . . . . . . . . . . 41
5.2.3 Longer Time Period . . . . . . . . . . . . . . . . . . . . . . 41
5.2.4 More Features to Explore . . . . . . . . . . . . . . . . . . . 41
A Chapter 1 of appendix 42
B Chapter 2 of appendix 46
Bibliography 49
vi
LIST OF TABLES
3.1 variables considered for the impact of commodity prices on agri-
cultural loan losses . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 Augmented dickey-fuller Test Statistics . . . . . . . . . . . . . . . 19
4.1 Goodness-of-fit for models . . . . . . . . . . . . . . . . . . . . . . 25
4.2 Charge-off ratio ARDL model result . . . . . . . . . . . . . . . . . 26
4.3 Provision ratio ARDL model result . . . . . . . . . . . . . . . . . 29
4.4 forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.5 Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
vii
LIST OF FIGURES
1.1 Commodity Price from USDA, 2010-2020 . . . . . . . . . . . . . . 1
1.2 Farmers’ Net Cash Income from USDA, 2010-2020 . . . . . . . . . 2
3.1 Chargeoff-Provision Correlation, 2010-2019 . . . . . . . . . . . . . 16
3.2 Farm Credit Services of America Areas . . . . . . . . . . . . . . . 18
3.3 Agricultural Index Seasonally Adjusted First Difference in Log
Value, 2010-2019 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.4 Crop Index Seasonally Adjusted First Difference in Log Value,
2010-2019 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.5 Livestock Index Seasonally Adjusted First Difference in Log
Value, 2010-2019 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.6 Poultry Index Seasonally Adjusted First Difference in Log Value,
2010-2019 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.7 Charge-off Ratio in Farm Credit Service of America, 2010-2019 . 22
3.8 Provision Ratio in Farm Credit Service of America, 2010-2019 . . 22
A.1 original dataset sample of commodity price . . . . . . . . . . . . 43
A.2 original dataset sample of loan loss features . . . . . . . . . . . . 43
A.3 Farm Credit Service of America, 2010-2019 . . . . . . . . . . . . . 44
A.4 Farm Credit Service of America forecast . . . . . . . . . . . . . . . 45
B.1 Augmented Dickey-Fuller Test Procedures . . . . . . . . . . . . . 47
B.2 ADF test results for provisions and chargeoffs . . . . . . . . . . . 47
B.3 ADF test results for indices . . . . . . . . . . . . . . . . . . . . . . 48
viii
CHAPTER 1
INTRODUCTION
Agricultural, crop and livestock commodity prices have been falling in the past
5 years(fig 1.1) in the United States. However, operating expenses have not
decreased. Farmers are seeing their profits squeezed as commodity prices fall
while operating expenses remain high. A downturn in gross farmers’ income is
shown in the past five years from the data provided by USDA (fig 1.2).
Figure 1.1: Commodity Price from USDA, 2010-2020
Further, U.S. and Chinese trade war this year, combined with bad weather
that has vexed U.S. farmers, has made some U.S. farmers reluctant to sell their
crops, constraining supplies for commodity traders. For the U.S. agricultural
sector, China’s tariffs on products, including soybeans, pork, and corn-based
1
fuel additive ethanol, have pressured already-low commodity prices. The recent
trade war with China has also exaggerated the low farmers’ income situation.
Figure 1.2: Farmers’ Net Cash Income from USDA, 2010-2020
If debtors are not able to meet their financial obligations, the credit risk of
lenders rises. Farmers experience difficulties in exporting essential farm goods
like soybeans because of the trade war and generating sufficient revenue due to
the low commodity prices; it has led to the author’s curiosity about the impact of
commodity prices on agricultural loans. The question is, how will commodity
prices fluctuation affect losses of loans lent to farmers? How long does it take
for the current low commodity price to affect the loans?
Farm Credit System (FCS) is a nationwide financial cooperative lending to
agricultural and aquatic producers, rural homeowners, agriculture-related and
rural businesses and cooperatives. It has a mission to support rural communi-
ties and agriculture with reliable, consistent credit and financial services. Farm
2
Credit raises funds in the world’s capital markets and put those funds to work
in rural America, augmenting commercial banks. The Farm Credit System helps
ensure the vibrancy of farms throughout rural America. With a nationwide net-
work of 72 customer-owned financial institutions, Farm Credit offers loans to
agricultural producers, farmer-owned cooperatives, and other agribusinesses.
It has a portfolio of loans of 272 billion dollars, of which 17% goes to cash grains,
9% goes to cattle as of December 31, 2018.
Farm Credit Administration is an independent, non-appropriated federal
agency in the executive branch of the Federal Government of the United States.
FCS institutions are federal oversight by the Farm Credit Administration (FCA).
Every Farm Credit System institution is required to prepare and file Uniform
Reports of Financial Condition and Performance (Call Report) with the FCA.
Principles for Call Reports require that institutions maintain a reasonable
and adequate allowance for loan losses, which approximately equals the current
amount of loans that will not be collected. It is also referred to as “provisions
for loan losses.” Losses on loans are to be charged off to the allowance when
known. LLP can reveal financial institutions’ forecast of the year’s net change
in possible loan losses.
In order to provide a narrower focus on individual financial institutions, this
study selected one particular institution with complete data in the past decade
from the 72 financial institutions in Farm Credit System. The data examined in
this study comes from Call Report on Farm Credit Administration website. The
provision and charge-off values of this institution in the past 10 years are used
to explore the relationship between provisions and chargeoffs and commodity
prices. An autoregressive distributed lagged model is applied to determine the
3
lagged effect of commodity price fluctuations.
The remainder of the paper is organized as follows. The first section sum-
marizes the literature on this particular topic. The second section introduces the
hypothesises and the basic empirical models. The next two sections present first
the data, and then results, including a forecast for the next six quarters. The last
section provides concluding remarks.
4
CHAPTER 2
LITERATURE REVIEW
The main focus of this section is to present some recent stylized facts concern-
ing agricultural-lending business, to set out the pertinent issues for analyzing
the relationship between commodity price fluctuations and agricultural loan
losses. Some of the general literature on loan loss provisions and loan write off
in financial institutions is also reviewed.
Keeton and Morris (1987) suggest that performance of agricultural and en-
ergy industries has a substantial influence on variation in loan losses. Low com-
modity prices are one of the three major reasons that led to the 1980’s farm
crisis (Featherstone and Brorsen 1994). Finance professionals are worried that
falling agricultural commodity prices depress farmers’ incomes and squeeze
their profit, which will eventually erode farmers’ ability to repay loans. (Jackie,
2015)
A series of articles about loan losses have discussed the motivations for fi-
nancial institutions to adjust loan loss provisions. The intention of setting loan-
loss provisions is that if a borrower does not repay the loans, the financial in-
stitutions can capture expected future losses. If the financial institution has set
aside more loan-loss allowance than the actual loan write-offs, the financial in-
stitution can absorb more unexpected losses without failing and imposing losses
on the Federal Deposit Insurance Corporation (for Commercial Banks) if all else
is held constant (Larry and Timothy, 2000). Carolyn and Shawn (1998) found
that loan loss provisions are used for capital management and are negatively
related to both future earnings changes and contemporaneous stock returns.
Setting loan loss reserves is also considered adequate as the method of account-
5
ing for loan losses involved the subtraction of loan losses from current income
or net worth when the loan was charged off, which is tax favorable to companies
(John, 1991).
A large number of studies has been conducted on non-performing loans and
its relationship with macroeconomics. Ricardas (2012) confirmed that doubt-
ful and non-performing loans in banks are highly depend on macroeconomic
changes in a country. Macroeconomic determinants include GDP, inflation, in-
terest rates, money supply, industrial production index, current account balance
and other indicators. Jonathan and Jefferey (2010) discover that the rate of in-
terest charged by the lender at origination and the borrower’s current ratio at
origination are the primary drivers of prepayment and default. They applied
survival analysis to examine factors influencing agricultural mortgage repay-
ment using a sample of farm credit system loans. Almir et al. (2019) studied the
impact of gross domestic product growth rate and inflation rate.
However, few studies have focused on the relationship between agricultural
loan losses and commodity price fluctuations, let alone quantifying the effects
of the latter. Agricultural commodity prices fluctuate from several factors in-
cluding weather conditions, natural disasters, global policies, and exchange rate
volatility. Agricultural commodity prices itself can explain macroeconomic fac-
tors vividly. Lord (1993) pointed out the fluctuations in agricultural commodity
prices can be interpreted by demand, supply and inventory. Accessing the im-
pact of agricultural commodity prices can be instrumental to the study of agri-
cultural loan losses. Burns et al. (2015) modeled the impact of lower commodity
prices on the credit risk in the agricultural industries. According to the survey
conducted by Iowa State University (2015), the majority of survey respondents
6
(94%) replied that lower commodity prices are a major negative factor on farm-
land values. Jiao Hua and Chang Jian (2013) analyzes that price fluctuations
of agricultural commodities have a significant impact on farmers’ income using
a VAR model. Johnson et al. (2017) summarized the modern literature focus-
ing on the probability of default on agricultural lending. He suggested that if
the implications of macroeconomic changes on loan default rates could be more
accurately measured, lenders could better forecast loan losses.
When modeling the loan losses, Almir et al. (2019) conclude that it is nec-
essary to use dynamic data panel models and different assessors to study non-
performing agricultural loans. They identified a statistically positive link be-
tween the levels of nonperforming loans from the previous and current period,
which explains that the conditions for granting loans and the target group have
not changed, as well as the habits of the borrowers. Dressler and Tauer (2016)
pointed out that most studies have focused on only one of the many components
when it comes to agricultural credit risk assessment. A partial least squares
(PLS) model is used to forecast bank losses, net charge-offs (James et al. 2018).
Swanepoel (2019) used auto-regressive distributed lag model for long-run US
household debt determinants. When modeling agricultural commodity prices,
Box and Jenkins (2015) proposed to use autoregressive integrated moving av-
erage (ARIMA) model to predict agricultural commodity prices. Rausser and
Carter (2017) also used ARIMA model to analyze the future prices of soybean,
soybean oil and soybean meal.
A common limitation in the literature on agricultural commodity prices and
loan losses is specifying the time it takes for the former to affect the latter. As
concluded above, the literature has discussed the causal effect of low commod-
7
ity prices on loan losses, but none has quantified the time it takes for commodity
price fluctuations to affect loan losses in the U.S. This article contributes to the
literature on this issue in three ways. First, by constructing a distributed lag
model including all the commodity price indices from USDA, the lagged ef-
fects of commodity price indices on loan loss features are confirmed. Second,
the time commodity price’s fluctuation takes to affect the loan market are calcu-
lated. Third, the future loan losses are forecasted based on the successful model.
8
CHAPTER 3
METHOD
3.1 Hypothesis
3.1.1 Charge-offs, Provisions and Commodities
Iftekhar and Larry (2004) demonstrate that net charge-offs are one of the most
significant factors determining the U.S. bank’s loan loss provisions. When
bankers make decisions on loan loss for the next year, they calculate the num-
ber based on the charge-off rates in the previous period. Loan loss charge-offs
are hypothesized as one of the factors that affect the variance of the loan loss
provisions.
Moreover, as shown in figure 3.1, loan loss provisions and loan loss charge-
offs have similar peaks and valleys with lagged correlation. It is possible that
loan loss provisions also have an impact on loan loss charge-offs.
Commodity price is another factor that is considered affecting loan loss fea-
tures as it is directly related to farmers’ income. Commodity price indices rep-
resent summarized prices of commodities in the same category. However, the
fluctuation in commodity prices might not result in a variance in loan loss fea-
tures in the current quarter. It is hypothesized that commodity price indices
impact loan loss features, but with certain lagged period.
9
3.1.2 Autoregressive
Behrens and Pederson (2007) found macroeconomic variables and autoregres-
sive impact of loans’ features useful when determining the probability of default
of a farm. They considered the autoregressive impact of the prior probability of
default, borrower-specific data, and macroeconomic variables. In our case, it is
hypothesized that loan loss provisions and loan loss charge-offs have autore-
gressice impact on themselves.
3.1.3 Lagged Effect
The third hypothesis is that loan loss features do not react only to the market
in the current quarter. For each kind of commodity, it possibly takes different
periods for loan loss features to respond to the change in commodity prices,
depending on the degree of influence of these commodities on farmers’ income.
3.2 ARDL Model
The autoregressive (AR) character of the time series model shows that the
present value of the variable depends on its past value. If the current variable
value is based on one preceding value, it is called first-order autoregressive, AR
(1), and if it is based on two preceding values, it is called second-order autore-
gressive, AR (2), and so on (Min and Guna, 2017). The time-series feature of
our variables determines that the AR part needs to be included in our model
to investigate the effect of past values on the present values. The provision ra-
10
tio and charge-off ratio are time series data spanning ten years. As discussed
in the literature review session, the provision ratio is affected by the previous
provision ratio because banks are using the previous ratio to make a current de-
cision. Charge-off ratio is also related to its previous ratio, as some loans might
be delayed for more than one quarter.
Apart from the previous values of the variables, the relationship between the
provision ratio and charge-off ratio is also worthy of study. As shown in Fig-
ure 3.1, the pattern of the charge-off ratio and provision ratio share the similar
trends with a certain lagged period. Additionally, the effect of commodity prices
on bank loans is intriguing. Hence, selecting a model that sufficiently includes
both loan features and commodities is necessary. After comprehensive research
on statistical modeling, the ARDL model is the best choice for this study.
An autoregressive distributed lag (ARDL) model is an ordinary least square
(OLS) based model with a mixed order of integration of dependent variables
and independent variables (Peseran and Shin, 1998). It captures dynamic effects
from both lagged dependent and independent variables. The two primary uses
of ARDL models are for forecasting and multiplier analysis.
The ARDL method has advantages over other approaches (Engle and
Granger, 1987; Johansen,1988; Johansen and Juselius, 1990). In our study, us-
ing ARDL has three key advantages. Firstly, the ARDL model can best explain
the lagged effects of both dependent and independent variables in our study. As
discussed in previous subsection, loan loss features have autoregressive impact
on themselves. Commodity price indices also have lagged effect on loan loss
features. A distributed lag model is more helpful as it includes the lagged ef-
fects of all independent variables and the autoregressive effect of the dependent
11
variable. Further, ARDL is desirable for forecasting loan loss features.
Secondly, serial correlation can be eliminated in the errors by including a
sufficient number of lags of the independent variable. ARDL model does not
require pre-testing order of integration of the variables.
Lastly, ARDL has the edge over another general forecasting model in inves-
tigating this question because it can be applied to studies that have a limited
sample size. Although the dataset spans ten years, the financial data were re-
ported quarterly, which only leaves us 40 observations in total.
3.3 Sample and Data Description
This section describes the data sources and presents main summary statistics.
3.3.1 Data Sources
The data in this research comes from two sources. The data source for commod-
ity price indices is from the United States Department of Agriculture National
Agricultural Statistics Service. The data is a sample of commodity prices re-
ceived for all agricultural products by month in the U.S. In total, there are 11
indices: Agriculture, Crop, Livestock, Fruit & Tree Nut, Vegetable & Melon,
Feed Grains, Food Grains, Oilseed, Dairy, Meat Animal, and Poultry & Egg.
The data is formatted as monthly data spanning a period from January 2010 to
August 2019.
The data source for loan loss features is from the quarterly call report data
12
from Farm Credit Administration. Call report data are provided in comma-
delimited text files, separated in quarters from March 2000 to September 2019.
The data files include three categories of financial data: consolidated report of
condition, consolidated report of income, regulatory capital framework. Ac-
cording to the Instructions for Preparing the Uniform Call Report Required by
the Farm Credit Administration, every farm credit system institution is required
under regulation, to prepare and file the Call Report with the FCA. Farm Credit
System institutions comprises three categories of financial institutions: A. Farm
Credit System Banks, B. Farm Credit System Associations, C. Farm Credit Sys-
tem service institutions. There are 79 financial institutions in total in the Farm
Credit Network.
3.3.2 Data Processing
Variable Definitions
Variables are shown in table 3.1.
Dependent Variable Formula
Loan Loss Provision Provisions/Total Loans*1000
Loan Loss Charge-off Charge-offs/Total Loans*1000
Independent Variable Formula
11 Commodity Price Indices Log(Commodity Price Indices/CPI)
Table 3.1: variables considered for the impact of commodity prices on agri-
cultural loan losses
Loan Loss Provision: Financial institutions set aside allocated funds at the
beginning of a period as a provision for uncollected loans. This provision is
used to cover several factors associated with potential loan losses, including
13
bad loans, customer defaults, and loan terms renegotiated that are less than
previously estimated.
Loan Loss Charge-off: When the creditor finds the loans to be ‘sent off as
non-collectible,’ they put the amount into the charge-off term in the financial
statement.
Commodity price index: It is a weighted index average of individual com-
modity prices. It represents the broad commodity asset class. There are 11
indices in general, which contain different types of Agricultural products like
Crop, Livestock, Fruit & Tree Nut, Vegetable & Melon, Feed Grains, Food
Grains, Oilseed, Dairy, Meat Animal, and Poultry & Egg.
Commodity Price Indices
Commodity price indices reflect the commodity market performance in selected
category. However, the commodity price is also influenced by other macroeco-
nomic factors. The indices data are dealt with by following four steps:
Firstly, for independent variables, the influence of inflation on commodity
prices over ten years are considered. We obtained CPI for All Urban Consumers
(CPI-U) series data from the Bureau of Labor Statistics. This CPI data includes
all items in U.S. city average, all urban consumers, not seasonally adjusted. The
series data number is CUUR0000SA0. The first month of the start year of the
time series data is set as 100. Commodity indices data is deflated by dividing
the CPI time series data. In the end, we obtain commodity price indices for 10
years without the affect from the inflation on the market.
14
Secondly, logarithm transformation is performed on the commodity price
data. Log transformation is instrumental when dealing with price data. Prices
in general tend to be growing exponentially over time and act on a multiplica-
tive scale. A log transformation realizes linearization of exponential growth. It
converts the multiplicative relationships to additive relationships and change
the trends from exponential to linear.
Thirdly, however, after plotting the data, it is evident that the data has a
strong seasonal effect. The time series data is decomposed, and the seasonal
components are removed. The deseasonalization of time series data is necessary
for understanding the underlying trends of commodity price data without the
influences of predictable seasonal patterns. Seasonal adjustment techniques are
based on the idea that a time series can be represented as the product of four
components: value of the long-term secular trend in series, value of seasonal
component, cyclical component, irregular component.
Lastly, to match the format of dependent variable data, the monthly com-
modity price indices in log value data are aggregated into quarterly presented
data.
Loan Loss Features
The RC1 Memoranda file and the RIE Analysis of Allowance for losses-loans,
notes, sales contracts, and leases file include loan loss features. There are 38
files in total for each type of file which span a period of from the first quarter in
2010 to the latest quarter in 2019. The 38 files are merged to make the data for-
mat in time sequence. Two files are generated after this process, both including
15
Figure 3.1: Chargeoff-Provision Correlation, 2010-2019
dotted:provisions, solid:charge-offs
columns - date, financial institution code, and variables.
The two tables (RIE and RC1) are joined together based on the same date and
same financial institution code. In reconciling date from data sets with the pro-
vided data, cleaning was necessary to ensure the same conventions were being
used (e.g., financial institutions code and date). Generally, for exploratory anal-
ysis and visualization purposes, incomplete and erroneous data were simply
omitted. Here, we have obtained a clean dataset with date, financial institution
code, charge-offs, provisions, total loans. Lastly, we computed the ratios as loan
loss provisions divided by total outstanding loans and loan charge offs divided
by total loans, scaled by 1000 times.
The provision ratio and chargeoff ratio are plotted in figure 3.1. Figure 3.1
16
shows that charge-offs and provision increase and decrease in similar trends.
This pattern has sparked the author’s interest to further look into the relation-
ship between the two.
Selecting Financial Institutions
After combing the data files, the list of top 10 financial institutions with the
most comprehensive data is generated. They are: First South Farm Credit, Farm
Credit of Northwest Florida, AgGeorgia Farm Credit, AgSouth Farm Credit,
GreenStone Farm Credit, Farm Credit Services of America, Northwest Farm
Credit Services, Cape Fear Farm Credit, AgCarolina Farm Credit.
Farm Credit Services of America is selected to study as it covers the agricul-
tural states of Iowa, Nebraska and South Dakota. According to the statistics re-
sults from USDA, Iowa ranked 2nd, Nebraska ranked 4th, South Dakota ranked
11th on the gross receipts of farms state ranking. These three states together
combine for 15.6 percentage points share of the total gross receipts of farms in
the U.S. Studying Farm Credit Services of America first has a great value for
explaining the farm finance in the U.S. Additionally, these states have less di-
versification of agricultural commodity production (Tauer, 2018). They produce
mainly grain and oil crops, livestock and livestock products, all of which are
sensitive to commodity prices. Furthermore, soybean and corn are important
exports, which are now affected by the U.S. China trade war.
Determining Stationary of Time Series
Before proceeding, the stationary of time-series data must be ensured.
17
Figure 3.2: Farm Credit Services of America Areas
The augmented dickey-fuller test is applied to the time-series data. Aug-
mented dickey-fuller test tests the null hypothesis that a unit root is present in
a time series sample. If the null hypothesis is not rejected, the series is non-
stationary.
The stationary of the loan loss features is tested. From figure 3.4, it appears
that the charge-off ratio is stationary. The augmented dickey-fuller test shows
the p-value is less than 0.01, which verifies that the charge-off ratio is stationary.
The same analysis procedure is applied to the provision ratio. The result shows
the p-value is 0.04, which indicates that the provision ratio is stationary on the
significant level of a 95 confidence interval.
The stationary of the monthly commodity price indices is tested. The first
18
test is applied to the monthly data after seasonal adjustment. The results of the
first test appear not stationary. The monthly data is differenced once, then tested
stationary again. The results of this time show all the commodity price indices
stationery.
The monthly commodity price indices first difference values are further ag-
gregated into quarterly data to match the format of loan loss features. The sta-
tionary of the quarterly commodity price indices is tested. The results point out
that all the commodity price indices are not stationary. The quarterly commod-
ity price indices are differenced once, then tested stationary again. This time the
results show the data is stationary.
After the thorough stationary process, as shown in table 3.2, all of our vari-
ables are stationary. More detailed procedures of calculations for making the
variables stationary can be found in the appendix chapter 2.
Variables adf test p-value
charge-offs ratio 0.01
provision ratio 0.04
crop index 0.01
poultry index 0.01
livestock index 0.01
agricultural index 0.01
Table 3.2: Augmented dickey-fuller Test Statistics
3.4 Modeling
The autoregressive distributed lag (ARDL) model is employed to explore the
relationship between independent and dependent variables. The ARDL model
19
Figure 3.3: Agricultural Index Seasonally Adjusted First Difference in Log
Value, 2010-2019
Figure 3.4: Crop Index Seasonally Adjusted First Difference in Log Value,
2010-2019
20
Figure 3.5: Livestock Index Seasonally Adjusted First Difference in Log
Value, 2010-2019
Figure 3.6: Poultry Index Seasonally Adjusted First Difference in Log
Value, 2010-2019
21
Figure 3.7: Charge-off Ratio in Farm Credit Service of America, 2010-2019
Figure 3.8: Provision Ratio in Farm Credit Service of America, 2010-2019
22
includes the AR component from the regression of y on lagged values of itself;
also, it includes the distributed lag effect of the x′i s.
Commodity price indices’ effect on the loan loss features are analyzed. The
dependent variables are the charge-off ratio or provision ratio of loans to farm-
ers in one particular financial institution from 2010 to 2019. Commodity price
indices first differences in log value and their lagged components are indepen-
dent variables.
When loan loss provisions is the dependent variable, loan loss charge-offs
and its lagged components are also independent variables, vice versa.
Formula 3.1 shows a general formula for this study. The following regres-
sions: formula 3.2 and formula 3.3 are estimated in this article, where CF repre-
sents charge-offs ratio from 2010 to 2019 in the financial institution, PV denotes
provision ratio, and IN denotes commodity price indices. The commodity price
indices can be more than one kind of commodity. The parameter alpha, beta,
gamma are the coefficients at different times for each variable.
BIC(Bayes information criterion) is the criterion for model selection in this
study. BIC has been commonly used for selecting the best model in time series
and linear regression. The lower BIC value indicates a better model fit. The BIC
value determines the number of optimal lags in the ARDL estimated model.
∑m ∑n
yt = δ + αiyt−i + β jxt− j + t (3.1)
i=1 j=0
∑m ∑n ∑k
CFt = δCF + αCFCFt−i + βCF PVt−i + γCF INt−i + t (3.2)
i=1 i=0 i=0
23
∑m ∑n ∑k
PVt = δPV + αPV PVt−i + βPVCFt−i + γPV INt−i + t (3.3)
i=1 i=0 i=0
24
CHAPTER 4
RESULTS AND FORECASTING
4.1 Results
The optimal lag length of models was calculated. Bayes information criterion
was used to determine the lag length. Goodness-of-fit of both models are re-
ported in table 4.1. Variables and their coefficients are reported in table 4.2 and
4.3. In the table, charge-offs represent the charge-off ratio at the financial in-
stitution, provision represents provision ratio at the financial institution, AG
denotes agricultural index, CP denotes crop index, LS denotes livestock index,
PT denotes poultry index. All of the indices are log values of quarterly first
differences.
R- squared Adj R-squared Number of obs Lags
Provision
0.95 0.77 33 ARDL(4,3,4,2,4,4)
charge-offs
0.66 0.42 33 ARDL(3,0,2,3,0,0)
Table 4.1: Goodness-of-fit for models
4.1.1 Charge-offs
Results of the charge-off ratio ARDL model are in table 4.2. The optimal num-
ber of lag components in the model, as determined by the BIC criterion, are
charge-offs lagged three periods, provision no lags; agricultural index lagged
two periods, crop index lagged three periods, livestock index and poultry index
none lagged.
25
Variables Coefficients t-Statistic P-value
charge-offs
L1 .296351 1.88 0.076
L2 -.1003577 -1.07 0.29
L3 .334492 4.42 0.000**
provision
L0 -.0442593 -0.98 0.338
AG
L1 -3.076075 -0.78 0.445
L2 -8.837892 -2.69 0.015*
CP
L1 10.16341 2.39 0.028*
L2 12.86014 3.52 0.002**
L3 4.573707 2.50 0.022*
LS
L0 4.26164 1.01 0.325
PT
L0 -.2250094 -0.21 0.834
constant
- .0429527 1.97 0.063
Table 4.2: Charge-off ratio ARDL model result
The R-squared of the model is 0.6550, which shows that this model can
explain 65.50 percent of the variance in the charge-off ratio. The adjusted R-
squared of the model is 0.42, which means after adjusting the effect of the num-
ber of included variables, this model can explain 42 percent of the variance in
the charge-off ratio.
One unit increase in the charge-off ratio three quarters ago leads to an in-
crease of 0.33 units in the current charge-off ratio. One unit increase in the
charge-off rate two quarters ago leads to a decrease of 0.1 units in the current
charge-off ratio. One unit increase in the charge-off rate in the last quarter leads
to an increase of 0.29 units in current charge-off ratio. One unit increase in the
provision rate in the current quarter leads to a decrease of 0.04 units in charge-
26
off ratio. A 1 percentage increase in the change in the agricultural index variable
in the last quarter causes a fall of 3.08 percentage in charge-off ratio currently.
A 1 percentage increase in the change in the agricultural index variable two
quarters ago causes a fall of 8.84 percentage in charge-off ratio currently. A 1
percentage increase in the change in the crop index variable in the last quar-
ter generates a 10.16 percentage increase in the current charge-off ratio. A 1
percentage increase in the change in the crop index variable two quarters ago
generates a 12.86 percentage increase in the current charge-off ratio. A 1 per-
centage increase in the change in in the crop index variable three quarters ago
generates a 4.57 percentage increase in the current charge-off ratio. Livestock
and poultry index variables do not have lagged effects on charge-off ratio. A 1
percentage increase in the change in the livestock index variable increases the
charge-off ratio by 4.26 percentage. A 1 percentage increase in the change in the
poultry index variable decreases the charge-off ratio by 0.23 percentage.
When looking at the coefficients of the independent variables, the p-value of
the third lag of charge-offs is significant, but not for the first and second lagged
period. This points out that the charge-off ratio three quarters ago has an influ-
ence on its current value. The p-value of the provision is not significant, which
confirms that the default possibility of agricultural loans has nothing to do with
the provision financial institutions set for themselves. This might conclude that
charge-offs come as a surprise to this financial institution, and they were not
able to accurately forecast loan losses and set aside loan loss provisions. The p-
value of the agricultural index is significant on the second lag but not the first.
This shows that the agricultural index starts to have an impact on loan perfor-
mance only from the last quarter. All three lag values of crop price index are
significant. This confirms that the crop price index plays a vital role in deter-
27
mining the performance of loans. This aligns with our hypothesis that the price
of crops has a determining effect on farmers’ capability to pay back loans. The
p-values of livestock and poultry are not significant, which shows that these
two commodities have a limited impact on loan performances.
The charge-off ratio has a positive relationship with its third lag. This in-
dicates that the increase of charge-offs in the past could increase charge-offs in
the current period. This agrees with our theory that some bad loans might stay
for more than one quarter. The agriculture index has a negative relationship
with charge-offs. This proves our hypothesis that when agricultural commod-
ity prices go up, there are fewer bad loans. Crop index has a positive relation-
ship with charge-offs. This does not align with our original belief that when
crop price goes up, fewer farmers default on their loans. However, it is possible
that the crop index starts to impact the loan losses negatively more than one
year ago, and the model eliminates that part when choosing the optimal model
bases upon the BIC criterion.
4.1.2 Provisions
Results of the charge-off ratio ARDL model are in table 4.3. While provision is
lagged four times, charge-off ratio is lagged three times, the agricultural index
is lagged four times, crop index is lagged two times, livestock index is lagged
four times, poultry index is lagged four times, we have the optimal number of
lag components in the model.
The R-squared of the model is 0.9574, which indicates that the model can ex-
plain 95.74 percent of the variance in the provision rate. The adjusted R-squared
28
Variables Coefficients t-Statistic P-value
provision
L1 .9553551 3.64 0.011*
L2 .3363163 1.73 0.135
L3 -.8677172 -3.73 0.010**
L4 .175769 0.98 0.366
charge-off
L0 -.1916185 -0.28 0.788
L1 .2169766 0.31 0.767
L2 -1.288333 -2.88 0.028*
L3 -.8195779 -1.87 0.110
AG
L0 12.2426 0.35 0.742
L1 338.9821 3.94 0.008**
L2 266.5442 4.04 0.007**
L3 8.765884 0.75 0.484
L4 25.58926 2.58 0.042*
CP
L0 36.06385 1.47 0.193
L1 -222.0728 -3.04 0.023*
L2 -227.9543 -4.05 0.007**
LS
L0 -7.770682 -0.36 0.734
L1 -143.8448 -3.50 0.01*
L2 -106.3202 -3.51 0.013*
L3 -24.22895 -1.97 0.096
L4 -85.80037 -4.41 0.005**
PT
L0 -7.239121 -1.80 0.123
L1 5.673904 0.76 0.474
L2 33.04491 3.42 0.014*
L3 31.16106 4.65 0.004**
L4 45.21302 4.37 0.005**
constant
- .2072602 2.65 0.038*
Table 4.3: Provision ratio ARDL model result
of the model is 0.77, which means after adjusting the effect of sampling size, this
model can explain 77 percent of the variance in the charge-off ratio.
29
One unit increases in provision ratio itself four quarters ago leads to an in-
crease in the provision rate of 0.18 units in the current quarter. One unit in-
creases in provision ratio three quarters ago leads to a decrease in the provision
rate of 0.87 units in the current quarter. One unit increases in provision ratio
leads to an increase in the provision rate of 0.34 units in the current quarter. One
unit increases in provision ratio in the last quarter leads to an increase in the pro-
vision rate of 0.96 units in the current quarter. One unit increases in charge-off
ratio three quarters ago leads to a decrease in the provision rate of 0.82 in the
current quarter. One unit increases in charge-off ratio two quarters ago leads
to a decrease in the provision rate of 1.29 units in the current quarter. One unit
increases in charge-off ratio in the last quarter leads to an increase in the pro-
vision rate of 0.22 units in the current quarter. One unit increases in charge-off
ratio in the current quarter leads to a decrease in the provision rate of 0.19 units
in the current quarter. A one percentage increase in the change in agricultural
index results in an increase in the provision rate of 12.24 percentage in the cur-
rent quarter. A one percentage increase in the change in agricultural index four
quarters ago results in an increase in the provision rate of 25.59 percentage in
the current quarter. A one percentage increase in the change in the agriculture
index variable three quarters ago results in an increase in the provision rate of
8.77 percentage in the current quarter. A one percentage increase in the change
in the agriculture index variable two quarters ago results in an increase in the
provision rate of 266.54 percentage in the current quarter. A one percentage in-
crease in the change in the agriculture index variable in the last quarter results
in an increase in the provision rate of 338.98 percentage in the current quarter.
A 1 percentage increase in the change in crop commodity price index variable
causes a rise in the provision rate of 36.06 percent in the current quarter. A 1
30
percentage increase in the change in crop commodity price index variable in
the last quarter causes a decrease in the provision rate of 222.07 percent in the
current quarter. A 1 percentage increase in the change in crop commodity price
variable two quarters ago causes a fall in the provision rate of 227.95 percent
in the current quarter. A 1 percentage increase in the change in livestock index
variable generates a fall in the provision rate of 7.77 percentage in the current
quarter. A 1 percentage increase in the change in livestock index variable in the
last quarter generates a fall in the provision rate of 143.84 percentage in the cur-
rent quarter. A 1 percentage increase in in the change in livestock index variable
two quarters ago generates a fall in the provision rate of 106.32 percentage in the
current quarter. A 1 percentage increase in in the change in livestock index vari-
able three quarters ago generates a fall in the provision rate of 24.23 percentage
in the current quarter. A 1 percentage increase in the change in livestock index
variable four quarters ago generates a fall in the provision rate of 85.80 percent-
age in the current quarter. A 1 percentage increase in the change in poultry and
egg index variable makes a fall in the provision rate of 7.24 percentage in the
current quarter. A 1 percentage increase in the change in poultry and egg index
variable in the last quarter makes a rise in the provision rate of 5.67 percentage
in the current quarter. A 1 percentage increase in the change in poultry and egg
index variable two quarters ago makes a rise in the provision rate of 33.04 per-
centage in the current quarter. A 1 percentage increase in the change in poultry
and egg index variable three quarters ago makes a rise in the provision rate of
31.16 percentage in the current quarter. A 1 percentage increase in the change in
poultry and egg index variable four quarters ago makes a rise in the provision
rate of 45.21 percentage in the current quarter.
The p-values of the first and third lag of provision is significant. This points
31
out that the provision ratio in the last quarter and three quarters ago have im-
pact on the current value. The p-value of the second lag of the charge-off ratio
is significant. This shows that the charge-off ratio three quarters ago has an im-
pact on the current provision rate. This indicates that it might take banks three
quarters to react to the charge-off ratio variance to make adjustments on the pro-
vision. The p-values of the first and second and fourth lags of the agricultural
index are significant. This demonstrates that it takes banks at least one quarter
to react to the change in agricultural commodity prices. The p-values of the first
and second lags of crop index are significant, which shows that the crop index
starts to influence banks’ decisions on setting provisions two quarters ago. This
shows that financial institutions also value the impact of crop index, so they
make adjustments according to the crop price promptly. The first, second, and
forth lags of livestock index are significant. The financial institution also consid-
ers livestock prices while making decisions on setting aside provisions. It takes
them one quarter or more to react to the market. As for the poultry index, the
second, third, and forth lags are significant. The price of poultry affects loan
provisions starting two quarters ago.
The agriculture index has a positive relationship with loan provisions. It
shows that when agricultural price increases, the bank raises the provision for
bad loans. This is opposite to our hypothesis but also interesting. The crop index
has a negative relationship with loan provisions. This proves our hypothesis
that when the crop price increases, the financial institutions will decrease the
loan provision because there will be a slighter possibility for farmers to default
on their loans. The livestock index variable has a negative relationship with
the provision ratio, too. We suppose it is due to the same reasoning with crop
index. However, the poultry index has a positive relationship with the provision
32
ratio. This indicates that when poultry price rises, the bank raises the provision
ratio. Lastly, the first lag of provision has a positive correlation with itself, and
the third lag of provision has a negative correlation with itself, which means
the bank is likely to keep consistent trends for two quarters when setting their
provisions. However, they will probably decrease their loan provisions after
three quarters.
For both models, the not significant p-value lags are not deleted as the fol-
lowing lag component is significant and the previous ones have to be included
when calculating the model. The selection of the model is based on BIC criteria.
The number of lag components are calculated to make BIC smallest. 1
4.2 Forecasting
4.2.1 Charge-offs
The commodity price and loan loss data we obtained from USDA and FCA
websites span a period from the first quarter of 2010 to the second quarter of
2019. Charge-off ratio is forecasted from the third quarter in 2019 to the last
quarter in 2020.
As shown in table 4.1, charge-off ratio is a function of provision ratio, com-
modity price indices and its own lags. To predict current charge-off ratio value
using the ARDL model in table 4.1, we need provision ratio value in the cur-
rent quarter, charge-off ratio values in the past three quarters, agricultural price
1Note: *(**) indicates significant level at 5 percent and 1 percent respectively
33
index values in the past two quarters, crop index values in the past three quar-
ters, livestock index value in the current quarter, and poultry index value in the
current quarter.
For example, to predict charge-off ratio in the third quarter of 2019, we need
provision ratio value in the third quarter of 2019. We also need livestock index
value and poultry index value in the third quarter of 2019. Hence, we need
to forecast the values of commodity price indices and provision ratio for six
quarters into the future to employ the ARDL model.
Provision ratio and commodity price indices are both time series price data.
There are several methods widely applied to forecast a time series, such as au-
toregressive integrated moving average model, exponential moving average,
and simple moving average. Here in this paper, we select simple moving aver-
age method to complete our forecasting.
Moving average is a commonly used technical method when analyzing time
series data. It helps smooth out price change by reducing random short-term
price fluctuations. Simple moving average is one of the two commonly used
moving average methods (simple moving average and exponential moving av-
erage). Simple moving average is calculated as formula 4.1.
Simple Moving Average method is used to calculate values for provision ra-
tio and commodity price indices from the third quarter in 2019 to the last quarter
in 2020. As shown in formula 4.1, simple moving average method is a division
of the sum of values in certain period over the number of time periods. Because
we are forecasting six quarter into the future, we determine the length of the
moving average to use is six quarters. N in formula 4.1 denotes number six. An
34
denotes value of either ratios or commodity price ratio in the Nth quarter.
After obtaining the values of commodity price indices and provision ratio
from the third quarter in 2019 to the last quarter in 2020, charge-off ratio values
in the next six quarters are forecasted using formula 3.2 and listed in table 4.1.
Table 4.4 shows the results of forecasting charge-off ratio in the next six quarters.
The values range from 0.6 to 0.8.
4.2.2 Provisions
Similarly, provision ratio in the next six quarters are forecasted using formula
3.3 and listed in table 4.2.
As shown in table 4.2, provision ratio is a function of charge-off ratio, com-
modity price indices and its own lags. To predict current provision ratio values
using the ARDL model, we need provision ratio values in the past four quarters,
charge-off ratio current value and values in the past three quarters, agricultural
price index values in the current quarter and past four quarters, crop index val-
ues in the current quarter and past two quarters, livestock index values in the
current quarter and past four quarters, poultry index values in the current quar-
ter and past four quarters.
For instance, to predict provision ratio in the third quarter of 2019, we need
provision ratios in the past four quarters, which can be found in the dataset we
created using data from Farm Credit Administration. We also need charge-off
ratio value, agricultural index value, crop index value, livestock index value
and poultry index value in the third quarter of 2019, which is not available in
35
the dataset used to estimate the ARDL model. Hence, we need to forecast the
values of commodity price indices and charge-off ratio for six quarters into the
future to apply the ARDL model.
Simple Moving Average method is again used to calculate values of charge-
off ratio from the third quarter in 2019 to the last quarter in 2020. Provision ratio
in the next six quarters are forecasted using formula 3.3 and listed in table 4.2.
The results are shown in table 4.4.
As shown in table 3.1, the ratio is calculated as the division of total charge-
off or provision number over total loans. Financial institutions can multiply the
ratio with the total loans from the institution to calculate the possible loan loss
charge-offs and provisions in the coming six quarters.
Date Provisions Charge-offs
9/1/19 0.24 0.07
12/1/19 0.37 0.06
3/1/20 0.29 0.07
6/1/20 0.33 0.08
9/1/20 0.21 0.07
12/1/20 0.21 0.07
Table 4.4: forecasting
A1 + AS MA 2
+ A3 + ... + An
=
n
(4.1)
36
4.2.3 Comparison to the Actual Number
The data that is used to generate the model covers a period from the first quarter
of 2010 to the second quarter of 2019. In table 4.5, the number of forecast provi-
sion and charge-off ratio is compared to the actual number in the third and last
quarter of 2019.
Date 9/1/19 12/1/19
Forecast Provisions 0.24 0.37
Actual Provisions 0.24 0.30
Forecast Charge-offs 0.07 0.06
Actual Charge-offs 0.09 0.06
Table 4.5: Comparison
The forecast results are not too far away from the actual number. The biggest
difference only has 0.07 difference in the number. The forecast provision ratio is
the same as the actual number in the third quarter of 2019. The forecast charge-
off ratio is the same as the actual number in the last quarter of 2019. It can be
concluded that the models work fine in short-term forecasting.
37
CHAPTER 5
CONCLUSION AND SUGGESTIONS
5.1 Conclusion
Using the commodity price data and loan feature data from Farm Credit Ser-
vices of America from 2010 to 2019, this study examines the relationship among
provision, charge-off ratio and commodity price in a financial institution in
Farm Credit System. Results showed that provision is dependent on four lags
of its own, three lags of charge-off ratio, four lags of the agricultural index, two
lags of crop index, four lags of livestock index, and four lags of poultry index.
Charge-off ratio is a function of 3 three lags of its own, provision, two lags of the
agricultural index, three lags of crop index, livestock index, and poultry index.
The model of provision explains a 77% variance in the response variable. The
model of charge-offs explains 42% variance in the outcome variable. We find ev-
idence of a significant relationship among loan loss charge-offs and agricultural
index, crop index. We also identify a significant correlation among loan loss
provision, loan loss charge-offs, agricultural index, crop index, livestock index,
and poultry index.
The results of our modeling supported our following hypothesis: firstly, we
hypothesize that in the four states that Farm Credit Services of America services,
crops appear to be the most critical production determining loan repayment;
thus, the crop price is critical to farmers’ income, as well as their ability to repay
the loans. Secondly, provision is a function of charge-offs as banks will make
decisions of provision based on charge-offs ratio in the past quarters. Thirdly,
commodity price impacts farmers’ ability to pay back loans. However, the im-
38
pact does not reflect on the number of provision and charge-offs immediately.
Moreover, the results of our model quantify the time it takes for commodity
prices to reflect on loan losses. According to the optimal models, the agricul-
tural index has an impact on loan loss charge-offs two quarters from now; the
crop index impacts loan loss charge-offs in the next quarter. The provision ra-
tio is influenced by the charge-off ratio three quarters before, agricultural index
one and two and four quarters before, crop index one and two quarters before,
livestock index one and two and four quarters before, poultry index two and
three and four quarters before.
Our study addresses a significant challenge associated with measuring the
provisions in the future considering the effect of commodity prices for financial
institutions, which give out loans mainly to agricultural lenders. The results of
this paper can be useful to management teams from financial institutions in the
Farm Credit System. Management teams could decide on setting future loan
provisions based on current commodity prices and the current charge-off ratio.
This article contributes significantly to the study of agricultural loans, Farm
Credit System, and commodities. There has not been literature quantifying the
lag effects of commodity prices on agricultural loans, which makes this article
valuable to its related readers.
5.2 Suggestions for Future Work
The author merged the financial data files for all 83 financial institutions in Farm
Credit System while conducting the research. The database now concludes all
39
the financial data from all statements for all 83 financial institutions from the
year 2010 to 2019. This database can further be used to test different hypotheses
regarding the Farm Credit System. Future work can conclude the following
perspectives:
5.2.1 Questions
There are the following questions that the author identified after comparing the
results of the model with the original guesses.
Firstly, as discussed in the result session, for model one, the coefficient of
crop index does not show the sign that we were expecting. We expected that the
crop index would have a negative relationship with charge-offs because when
crop index goes down, it is reasonable that more charge-offs will happen be-
cause farmers would have gained less income.
Secondly, the agricultural index and poultry index surprisingly have posi-
tive relationships with loan provisions. We would guess that the relationship is
negative because when the indices go down, the bankers might raise the provi-
sion due to their prediction of more loan losses. However, the coefficients in the
provision model is harder to tell because the provision is subjective prediction
from financial institutions that could be impacted by many other factors.
The author hopes these questions can be resolved after examining more
cases in the Farm Credit System. However, if the same questions are identi-
fied when studying other cases, it is reasonable to include other hypotheses to
test the data—for instance, the effect of taxation on loan provisions.
40
5.2.2 More Cases from Farm Credit System
Due to the time limit in the author’s master’s degree, only one out of 83 finan-
cial institutions is examined. Future researchers are welcome to further test the
conclusion of this article by using data for other financial institutions from the
database.
5.2.3 Longer Time Period
When trying to find the data for financial institutions, the author finds that the
call reports on government website only starts from 2010. The author has also
tried to obtain data directly from Farm Credit Administration, but they were
unable to provide data that are not public on the website. The earlier data is not
available electronically. This has resulted in a small sample as all the financial
data are quarterly data. Small sample with too many features might cause the
model over fitting. Future researchers can re-test this model when having data
for more years in the future.
5.2.4 More Features to Explore
This article mainly discusses the effects of commodity price on loan features.
However, as we discussed above in the literature review session, there are other
factors that might affect loans, such as the tax reasons, the general macroeco-
nomics. Future study can also conclude more features in the model when the
database is enlarged.
41
APPENDIX A
CHAPTER 1 OF APPENDIX
Links to data:
1. https://www.fca.gov/bank-oversight/call-report-data-for-download
2. https://www.nass.usda.gov/Charts and Maps/Agricultural Prices/index.php
42
Figure A.1: original dataset sample of commodity price
commodity price indices data file from USDA website
Figure A.2: original dataset sample of loan loss features
loan feature data file from Farm Credit Administration Website, UNINUM is financial
institution code, PROVLNS is loan provisions, CHGOFFLNS is loan charge-off
43
Figure A.3: Farm Credit Service of America, 2010-2019
the dataset used to construct models in this study
44
Figure A.4: Farm Credit Service of America forecast
the dataset used to forecast, green number in AG, CP, LS, PT columns are the forecast
results of the indices using simple moving average method, red number in provision
and chargeoff columns are the forecasting results of the ratios using simple moving
average method, green number in provision and chargeoff columns are the forecasting
results of the ratios using ARDL model
45
APPENDIX B
CHAPTER 2 OF APPENDIX
Chapter 2 of appendix explains Augmented Dickey-Fuller Test procedures on
the original data set and their results.
Taking 6 monthly agriculture price index values for example here, we show
the calculations that have been done to generate the final variable AG: We let
AG1, Ag2, AG3, AG4, AG5, AG6 denote agriculture index log values in period
1,2,3,4,5,6.
Firstly, we conduct Augmented Dickey-Fuller Test on the time series data
A1-A6. The result shows not stationary.
Secondly, we take the first difference of AG1-AG6, which gave us 6 new
values. Let AG’1-6 denote them. The results are: AG’1 = 0, AG’2 = AG2-AG1,
AG’3 = AG3-AG2, AG’4 = AG4- AG3, AG’5 = AG5- AG4, AG’6 = AG6-AG5
Thirdly, we aggregate AG’1-6 into quarterly data. The method it to take the
average of every three values. We let qAG1,qAG2 denote the results. qAG1 =
(AG’1+AG’2+AG’3)/3, qAG2 = (AG’4+AG’5+AG’6)/3
Forthly, we test the stationary of the quarterly time series data again. The
result shows not stationry. Thus, we take the first differnce of qAG1-2 time
series data. Let qAG’1, qAG’2 denote the results. qAG’1 = 0, qAG’2= qAG2-
qAG1
Lastly, apply ADF test on qAG’1-2, the results show stationary. qAG’1-2 are
the final agriculture index variable used in the ARDL model.
46
Figure B.1: Augmented Dickey-Fuller Test Procedures
Figure B.2: ADF test results for provisions and chargeoffs
The procedures can be different. However, to minimize the effect of differ-
encing on the data, the monthly data is first made into stationary before aggre-
gated into quarterly data.
47
Figure B.3: ADF test results for indices
48
BIBLIOGRAPHY
1. Barth, J., Joo, S., Kim, H., Lee, K. B., Maglic, S., Shen, X., Bank, R.
(2018). Forecasting net charge-off rates of banks: A PLS approach. Unpublished
Manuscript.
2. Behrens, A., Pederson, G. D. (2007). An analysis of credit risk migration pat-
terns of agricultural loans. Agricultural Finance Review, 67(1), 87–98.
3. Burns, C., Tulman, S., Harris, J. M. (2015). Farm financial stress in a changing
economic environment: simulating credit risk with new imputed ARMS data
on farm debt (No. 330-2016-13624).
4. Carter, C., Rausser, G., Smith, A. (2017). Commodity storage and the market
effects of biofuel policies.
5. Dielman, T. E. (1986). The theory and practice of econometrics. International
Journal of Forecasting, 2(2), 245–246.
6. Dressler, J. B., Stokes, J. R. (2010). Survival analysis and mortgage termina-
tion at AgChoice ACA. Agricultural Finance Review, 70(1), 21–36.
7. Dressler, J. B., Tauer, L. W. (2016). Estimating Expected and Unexpected
Losses for Agricultural Mortgage Portfolios. American Journal of Agricultural
Economics, 98(5), 1470–1485.
8. FARM Credit, instruction for preparing call report
9. Hasan, I., Wall, L. D. (2004). Determinants of the loan loss allowance: Some
cross-country comparisons. Financial review, 39(1), 129-152.
10. Johnson, A. M., Boehlje, M. D., Gunderson, M. A. (2017). Agricultural credit
risk and the macroeconomy. Agricultural Finance Review.
11. Keeton, W. R., Morris, C. S. (1987). Why do banks’ loan losses differ. Eco-
nomic review, 72(5), 3-21.
49
12. Krishna, K. (1993). Imperfect competition and international commodity
trade: Theory, dynamics, and policy modelling. Journal of Economic Litera-
ture, 31(4), 1983–1985.
13. Liu Jiao-hua, Li Chang-jian. (2013). The Effect of Agricultural Commodities
Prices Increase on Farmers’ Income – An Empirical Study Based on VAR Model.
International Management Review, 9(1), 42–47.
14. Mileris, R. (2012). Macroeconomic determinants of loan portfolio credit risk
in banks. Inžinerinė ekonomika, 496-504.
15. Muhović, A., Radivojević, N., Ćurčić, N. (2019). Research of factors of
non-performing agricultural loans by primary data panels. Economics of Agri-
culture, 66(2), 569-578.
16. Pesaran, M. H., Shin, Y. (1998). An autoregressive distributed-lag modelling
approach to cointegration analysis. Econometric Society Monographs, 31, 371-
413.
17. Shrestha, M. B., Bhatta, G. R. (2018). Selecting appropriate methodological
framework for time series data analysis. The Journal of Finance and Data Sci-
ence, 4(2), 71–89.
18. Stewart, J. (2015). Ag Lenders Brace for Impact of Falling Commodity Prices.
American Banker, (F313).
19. Swanepoel, E. (2019). Auto-Regressive Distributed Lag Model for Long-Run
US Household Debt Determinants. Ezelda Swanepoel (2019). Auto-regressive
Distributed Lag Model for long-run US household debt determinants. Invest-
ment Management and Financial Innovations, 16(3), 40-48.
20. Tauer, L. W. (1992). Diversification of production agriculture across individ-
ual states. Journal of Production Agriculture, 5(2), 210-NP.
21. WALLAND, L., Koch, T. (2000). Bank Loan-Loss Accounting: A Review of
50
Theoretical and EmpiricalEvidence. Economic Review. Federal Reserve Bank of
Atlanta.
22. Walter, J. R. (1991). Loan loss reserves. FRB Richmond Economic Review,
77(4), 20-30.
23. Wilson, G. T. (2016). Time Series Analysis: Forecasting and Control, 5th Edi-
tion, by George E. P. Box, Gwilym M. Jenkins, Gregory C. Reinsel and Greta M.
Ljung, 2015. Published by John Wiley and Sons Inc., Hoboken, New Jersey, pp.
712. ISBN: 978-1-118-67502-1. Journal of Time Series Analysis, (5), 709.
51