Satya Dhavala Derek Smith Thesis

Using Big Data for Decisions in Agricultural Supply Chain
by
Derik Lafayette Smith
B.S. Automotive Technology Management and Supply Chain
Brigham Young University-Idaho, 2010
and
Satya Prakash Dhavala
Master of Business Administration
Indian Institute of Science, 2005
B.Tech. Electronics and Communications

Jawaharlal Nehru Technological University, 2002
Submitted to the Engineering Systems Division

in partial fulfillment of the requirements for the degree of
Master of Engineering in Logistics

at the
Massachusetts Institute of Technology

June 2013
© 2013 Derik Smith and Satya Dhavala. All rights reserved.
The authors hereby grant to MIT permission to reproduce and to distribute publicly paper and
electronic copies of this document in whole or in part.
Signature of Authors………………………………………………………………..……………….
Master of Engineering in Logistics Program, Engineering Systems Division
May 10, 2013
Certified by………………………………………………………………………………………….
Dr. Bruce C. Arntzen
Executive Director, Supply Chain Master’s Program,
MIT Center for Transportation and Logistics
Thesis Supervisor
Accepted by…………………………………………………………………………………………
Prof. Yossi Sheffi
Elisha Gray II Professor of Engineering Systems, MIT
Director, MIT Center for Transportation and Logistics
Professor, Civil and Environmental Engineering, MIT
1
Using Big Data for Decisions in Agricultural Supply Chain
by
Derik Smith and Satya Dhavala
Submitted to the Engineering Systems Division

on May 10, 2013 in Partial Fulfillment of the
Requirements for the Degree of Master of Engineering in Logistics
Abstract
Agriculture is an industry where historical and current data abound. This paper investigates the
numerous data sources available in the agricultural field and analyzes them for usage in supply
chain improvement. We identified certain applicable data and investigated methods of using
this data to make better supply chain decisions within the agricultural chemical distribution
chain. We identified a specific product, AgChem, for this study. AgChem, like many agricultural
chemicals, is forecasted and produced months in advance of a very short sales window. With
improved demand forecasting based on abundantly-available data, Dow AgroSciences, the
manufacturer of AgChem, can make better production and distribution decisions. We analyzed
various data to identify factors that influence AgChem sales. Many of these factors relate to
corn production since AgChem is generally used with corn crops. Using regression models, we
identified leading indicators that assist to forecast future demand of the product.
We developed three regressions models to forecast demand on various horizons. The first
model identified that the price of corn and price of fertilizer affect the annual, nation-wide
demand for the product. The second model explains expected geographic distribution of this
annual demand. It shows that the number of retailers in an area is correlated to the total
annual demand in that area. The model also quantifies the relationship between the sales in
the first few weeks of the season, and the total sales for the season. And the third model
serves as a short-term, demand-sensing tool to predict the timing of the demand within certain
geographies. We found that weather conditions and the timing of harvest affect when AgChem
sales occur.
With these models, Dow AgroSciences has a better understanding of how external factors
influence the sale of AgChem. With this new understanding, they can make better decisions
about the distribution of the product and position inventory in a timely manner at the source of
demand.
Thesis Supervisor: Dr. Bruce C. Arntzen

Title: Executive Director, Supply Chain Master’s Program, Center for Transportation and Logistics
2
Acknowledgements
We both thank Dr. Bruce Arntzen and Josh Merrill (Dow Agro Sciences) for their valuable inputs
throughout this effort. We also thank Thea Singer for assisting us in our writing efforts.
We also thank our respective families for their support, encouragement, and love during this
period.
3
Table of Contents
List of figures ................................................................................................................................... 6
List of tables .................................................................................................................................... 7
1 Introduction ............................................................................................................................ 8
1.1 Agricultural Data .............................................................................................................. 8
1.1.1 Applications of this Data ........................................................................................... 9
1.2 Agricultural Supply Chain ................................................................................................. 9
1.2.1 Product Flow through the Agricultural Chemical Supply Chain ............................. 10
1.2.2 Information Flow through the Supply Chain .......................................................... 11
1.2.3 Our Sponsor Company ............................................................................................ 11
1.3 Problem Statement ........................................................................................................ 12
1.3.1 AgChem – The Selected Product............................................................................. 12
1.4 Thesis Outline ................................................................................................................. 13
2 Literature Review .................................................................................................................. 14
2.1 What is Big Data? ........................................................................................................... 14
2.2 Forecasting in Agriculture: Review of Fertilizer Forecasting Studies ............................. 15
2.3 Demand Sensing and Short-term demand forecasting.................................................. 18
2.4 Chapter Summary........................................................................................................... 18
3 Methodology......................................................................................................................... 19
3.1 Understanding the Factors influencing Demand ........................................................... 19
3.2 Dividing the Problem ...................................................................................................... 22
3.3 Model Development: OLS Regression ............................................................................ 23
4 Data and Analysis .................................................................................................................. 25
4.1 Data Summary ................................................................................................................ 25
4.2 Analysis and Results ....................................................................................................... 26
4.3 Preliminary Analysis ....................................................................................................... 26
4.4 Model-1: Annual Sales Model ........................................................................................ 28
4.5 Model-2: Geographical Sales Model .............................................................................. 32
4.6 Model-3: Short-term Demand Model ............................................................................ 35
4.6.1 Timing of Product Application ................................................................................ 35
4.6.2 Harvest date as a trigger point for fall application ................................................. 37
4.6.3 Temperature and precipitation as predictors of weekly sales ............................... 38
5 Conclusion ............................................................................................................................. 45
4
5.1 Models and Significant Variables ................................................................................... 45
5.2 Implied Benefits to DAS.................................................................................................. 46
5.3 Next Steps and Additional Research .............................................................................. 46
5.4 Considerations for Future Applications ......................................................................... 47
5.4.1 Data Quality ............................................................................................................ 47
5.4.2 Application City not Retailer City ............................................................................ 48
5.4.3 EDI for Multiple Years ............................................................................................. 49
Appendix ....................................................................................................................................... 50
References .................................................................................................................................... 53
5
List of figures
Figure 1 Agricultural chemical product flow................................................................................. 10
Figure 2 Agricultural chemical information flow .......................................................................... 11
Figure 3 Increasing demand over the last 5 years ........................................................................ 13
Figure 4 Factors influencing demand of the product ................................................................... 20
Figure 5 Increasing demand and seasonality of AgChem observed over a 10 year period ......... 27
Figure 6 Product flow between echelons ..................................................................................... 28
Figure 7 Annuals sales trending along with fertilizer and corn price ........................................... 29
Figure 8 Geographic plot of fall and spring quantities shipped (2007) ........................................ 36
Figure 9 Peak sales occur within 1-2 weeks after harvest completion ........................................ 37
Figure 10 Weekly sales correlate with temperature for Ames, IA ............................................... 40
Figure 11 Peak sales occur in spring for Avon, IN ......................................................................... 43
Figure 12 Geographical plot of fall and spring quantities shipped (2008-2009) .......................... 50
Figure 13 Geographical plot of fall and spring quantities shipped (2009-2010) ......................... 51
6
List of tables
Table 1 Factors influencing the demand and their granularity .................................................... 21
Table 2 Correlation coefficient matrix for all variables ................................................................ 30
Table 3 Regression results of the annual model ........................................................................... 30
Table 4 Regression results of the annual model with only significant factors ............................. 31
Table 5 Summary of factors for geographic model ...................................................................... 33
Table 6 Regression results of geographic model .......................................................................... 34
Table 7 Regression results of geographic model with season start sales included ...................... 34
Table 8 Weekly quantity sold data for Ames, IA........................................................................... 39
Table 9 Regression results of weekly model using entire season data ........................................ 40
Table 10 Regression results of weekly model using only fall data ............................................... 41
Table 11 Weekly quantity sold data for Avon, IN ......................................................................... 42
Table 12 Regression results of weekly model using entire season data ...................................... 43
Table 13 Regression results of weekly model using only spring data .......................................... 44
Table 14 Models and variables ..................................................................................................... 46
7
1 Introduction
Our thesis investigates how to make better supply chain decisions using the rich data available
in the agricultural industry. We begin with an analysis of various types of agricultural data and
the agricultural supply chain. We then describe the role of our sponsor company in the
agriculture distribution chain and present the problem they face. Subsequently, we explain the
process followed to analyze the problem and identify solutions.
1.1 Agricultural Data
As the world’s computing capability exponentially increases, our ability to gather, maintain, and
analyze data also increases dramatically. Data is more abundant in nearly every industry and
academic realm. The agricultural industry is no different. In some senses, there is more data in
the agricultural industry than in most others. Agriculture is one of the world’s oldest
occupations and trades. For millennia, farming practices have been passed on and improved
upon. For centuries, agricultural data has been collected, tracked, and analyzed.
Agricultural data ranges from crop yields in certain geographic areas to erosion, weather, and
climate trends. This data is collected, analyzed, and in most cases, shared by a variety of
sources: private companies, governmental agencies, and research universities. Some of this
data is published in journals, available through web-based databases or sold by the data
collection sources. This data varies in scope and detail; some data sources are very granular
while others are more general.
In companies that produce agricultural products, employee teams of R&D engineers and
researchers gather data and monitor trends to improve seeds, herbicides, pesticides, fertilizers,
and agricultural equipment. The information and knowledge gained by these companies is
generally patented, guarded, and used competitively within the marketplace. However, the
practices of data sharing and knowledge transfer are becoming more widely accepted within
some trade organizations and supplier/customer partnerships.
8
Additionally, government agencies and bureaus such as the United States Department of
Agriculture (USDA), amass large amounts of data. The USDA consists of 17 agencies, each
responsible for monitoring and assisting a different facet of the US agriculture industry. Some
of these agencies include the Agricultural Research Service (ARS), the Farm Service Agency
(FSA), and the National Agricultural Statistics Service (NASS). Each of these agencies gathers
and tracks masses of data which relate to its specific responsibilities. Most of this data and
research is published in reports and is publically available from the USDA or on its website.
Universities and other research centers also add to the overwhelming amount of agricultural
research and data. Land Grant universities have been called “the state-based component of the
public agricultural research system” (National Research Council, 1995). These universities
research a variety subjects in an effort to advance the knowledge and practices within the
agricultural field.
1.1.1 Applications of this Data
Once data is collected, it is used by a variety of people and organizations to improve the
effectiveness of today’s agricultural industry. Individual growers use crop, weather, and soil
data to make decisions about their upcoming growing season and long-term sustainability of
their farms. Companies who supply seed, chemicals, and other farming necessities monitor this
data to create demand forecasts and production plans. Crop insurance companies monitor
factors such as climate, yield, and farm economic data to constantly fine-tune their policies.
Much of this data is widely used throughout the industry. However, there are some who feel
that better operational and supply chain decisions could be made through careful compilation
and analysis of particular segments of this data. Our sponsor company feels this way.
1.2 Agricultural Supply Chain
When people hear the word “agriculture,” they most likely think of large fields of corn in Iowa,
citrus groves in Florida, or apple orchards in Washington. Each of these scenarios is at the
grower level. However, many other players are involved in the value chain of these vital
9
products. They include producers of seeds, fertilizers, chemicals, and agricultural equipment as
well as the retail organizations which market and sell these products.
Our principal interest is in the challenges faced by agricultural chemical producers. In this
section, we will elaborate on those parts of the supply chain which are relevant to this sector of
the agricultural industry.
1.2.1 Product Flow through the Agricultural Chemical Supply Chain
The supply chain for agricultural chemicals is large and complex. Manufacturers purchase
various components from a multi-echelon supplier base. Active and non-active ingredients are
combined in specific manufacturing sites. Often, one site will manufacture a chemical for
distribution globally. Manufacturing sites often produce a variety of chemicals and must plan
production based on forecasts and seasonality of demand. If the product is sold in bulk, it may
be shipped to a manufacturer-owned terminal where it is stored until it is shipped to a
distributor. If the product is to be sold in smaller quantities, it is sent to a re-packer who
packages the chemical in 250 gallon totes, 2.5 gallon jugs, or other suitable retail packs for
liquid applications and bags or envelopes for dry applications. After being packed, the product
is sent to a distributor.
Most manufacturers do not sell directly to growers or retailers, but instead sell through major
distributors. These distributors market products to retailers across the country. Retailers hold
bulk or packaged product and sell it to individual growers for application on fields and crops.
Figure 1 gives a representation of the product flow through this supply chain.
Manufacturer Distributor Retailer Grower
Figure 1 Agricultural chemical product flow
10
1.2.2 Information Flow through the Supply Chain
Traditionally, manufacturers have only received information about their sales and shipments to
distributors. Distributors have been very protective of information about their sales and
customers because they have feared that the manufacturers could circumvent them and sell
directly to the retailers. However, in recent years, due to manufacturers’ request and
investment in technologies such as electronic data interchange (EDI), distributors and retailers
have been sharing information with the manufacturers in an effort to improve supply
availability and reduce distribution costs. This information, represented in Figure 2, includes
inventory levels at distributor facilities, point-of-sale (POS) data from retailers, and grower
demographics.
Manufacturer Distributor Retailer Grower
Inventory Levels POS Data Demographics
Figure 2 Agricultural chemical information flow
1.2.3 Our Sponsor Company
Dow AgroSciences (DAS) is the sponsor of our project. On their website, they explain their
background in the following way:
We began as the plant sciences business of The Dow Chemical Company, which was
founded in 1897. A joint venture with Eli Lilly and Company in 1989 created DowElanco.
With a focus on sustainable agriculture, DowElanco combined the leading chemistries of
The Dow Chemical Company with those of the agricultural division of Eli Lilly. In 1997,
The Dow Chemical Company acquired full ownership of the business and named the
business Dow AgroSciences. Today, we employ more than 7,700 people worldwide, and
our 2012 global sales were $6.4 billion (U.S.) (About Us).
11
DAS offers a variety of products to the agricultural industry. These products include
“insecticides, herbicides, fungicides, fumigants, nitrogen stabilizers, seeds, traits, and oils”
(Products).
1.3 Problem Statement
We intend to identify and analyze data which can be used by DAS to understand the changes in
demand for a specific product and make better-informed decisions throughout their supply
chain. For this analysis, we have selected a product called AgChem. AgChem is primarily used
on corn, thus, we will focus on those factors associated with its production, such as corn prices
and climate. Our intention is to identify potential influences on the demand of AgChem among
the profusion of agricultural data.
1.3.1 AgChem – The Selected Product
AgChem is an agricultural chemical and is manufactured in two formulations, each under a

separate name. AgChem is generally used with the production of corn, and has been produced
and sold by the company for many years. This product is available in two varieties – AgChem
V1 and AgChem V2, with V1 being applied predominantly during fall and V2 during spring
season.
Demand for AgChem had stabilized and was believed to have matured. However, in recent
years, the demand has grown significantly. Figure 3 shows the monthly sales over the past five
years. As we can see there is a sudden and steep increase in sales. The company is interested in
knowing what factors have influenced this trend and what to expect for future demand.
12
Figure 3 Increasing demand over the last 5 years
The market for agricultural chemicals and fertilizers is seasonal and has a long production and
distribution lead times. Most of these products, including AgChem, are only applied during very
specific periods in a crop’s lifecycle. The demand for each product can vary significantly from
year to year depending on the mix of crops planted, weather and climatic conditions, farm
economics, as well as many other factors. Thus, companies like DAS must forecast and produce
chemicals months in advance for a short period of uncertain demand.
1.4 Thesis Outline
This section presented agricultural data, analyzed the agricultural chemical supply chain,
introduced our sponsor company, and stated the problem we have sought to solve. Next, in
our Literature Review, we will expand on terminology pertinent to this study and cover some
research that has been conducted with big data and agriculture forecasting. In Chapter 3,
Methodology we describe the analytical techniques we used to evaluate factors driving the
demand. Chapter 4 presents the data, our analysis, the models that were developed and the
results. We conclude with Chapter 5 where we discuss the results and their significance to our
sponsor company.
13
2 Literature Review
One of our objectives is to use the assortment of data available in the agriculture domain to
develop a short-term, geographic forecast of product consumption. In order to analyze this
problem better, we focus our literature review on three aspects: 1) big data and big data in
agriculture, 2) forecasting in the agriculture and farming sector, and finally 3) demand sensing
and short-term demand forecasting. While we were unable to find any literature that exactly
addresses our thesis problem, we aim to understand the concepts surrounding big data and
learn from data driven decision processes in agriculture.
2.1 What is Big Data?
While a number of definitions exist, analyst firm Gartner defines big data as “high-volume, high-
velocity and high-variety information assets that demand cost-effective, innovative forms of
information processing for enhanced insight and decision making.” Brynjolfsson, Hitt, and Kim
(2011) conducted empirical research and concluded that organizations’ performance is directly
related to their ability to make data driven decisions. This makes it important to understand the
success factors required to adopt big data techniques in organizations.
McAfee and Brynjolfsson (2012) postulate that three characteristics, volume, velocity, and
variety, make the field of big data fundamentally different from analytics and hence companies
require a different set of strategies to become big data enabled. In addition to the technical
challenges, big data also presents a set of managerial challenges. Most important among these
is the organization’s newfound ability to make data-driven decisions and hence move away
from intuition-based decision making. Leadership, talent management, technology, decision
making, and company culture are critical factors to be successful in leveraging big data for
organizational performance (McAfee, 2012).
On the agricultural side, Ashlee Vance (2012) of Business Week reports of innovative company,
Climate Corp which uses localized weather data to predict crop yields and uses this data to
tailor crop insurance premiums to individual farmers. Climate Corp stores and analyzes more
14
than 60 years of weather data consisting of precipitation, temperature, and soil conditions with
a granularity of 2 1/2 mile radius in order to arrive at the crop yield forecasts (Vance, 2012).
2.2 Forecasting in Agriculture: Review of Fertilizer Forecasting Studies
Since agriculture is tightly linked to the food-security of a nation, a lot of research has been
conducted by governmental agencies to ensure a sustainable and productive agricultural
future. In the United States, land grant universities are also a major source of research. They
have been the suppliers of yearly growth forecasts such as crop yields and prices as well as
short-term outlooks on crop and input prices. The majority of forecasting in agriculture uses
econometric modeling. This entails using exogenous variables to forecast crop yields, fertilizer
consumption, etc. We are particularly interested in research related to fertilizer consumption
forecasting since it relates somewhat to the use of AgChem.
The earliest known work in fertilizer forecasting was performed by Vail (1927), and then by
Mehring and Shaw (1944). They tried to investigate the relationship between nitrogen
consumption per acre with crop value, soil nitrogen content, and proportion of cash generated
from livestock. They all investigated the relationship between expenditure on fertilizer with
lagged farm income. Mehring and Shaw concluded that farmers always tend to spend a
constant proportion of their income on fertilizer. Vail did not find any significant relationship
between fertilizer consumption and fertilizer price.
Zvi Griliches (1958) completed another study of fertilizer consumption with the primary
objective of estimating the short-term and long-term demand elasticity. He found that
technological changes that resulted in new production technologies led to a reduction in the
real prices of fertilizer, thus leading to large scale adoption and increased use of fertilizer.
Griliches’ hypothesis was that an increase in fertilizer use was predominantly driven by a
decrease in the price of fertilizer in relation to other farm inputs and the price received for
output. He modeled fertilizer consumption in the current season as a function of the fertilizer
price and the previous season’s fertilizer consumption using the data from 1911-1956. Instead
of using the weight of fertilizer consumed (i.e. total tons consumed per year) he used the USDA
15
measure of “principal plant nutrients”, which measures the total quantities of the three main
components of the fertilizer – potassium, nitrogen and phosphate. The results he obtained
indicated a very strong relationship between the price of fertilizer and the consumption (R2 ≈
0.98) (Griliches, 1958). This study helps us understand the role price plays in factor selection of
farmers.
The Food and Agriculture Organization of the United Nations (FAO), which is based in Rome,
Italy, publishes a global outlook on fertilizer demand and consumption and other agricultural
trends such as crop yields and grain production. Parthasarathy of FAO, in “Demand Forecasting
for Fertilizer Marketing” (1994) analyzed various factors that influence fertilizer consumption
along with long-term and short-term forecasting models for the same. In addition to exogenous
factors such as farmers’ purchasing power, which is determined by the affordability of fertilizer,
and farmers’ cash liquidity, availability of the fertilizer also influences the consumption.
Infrastructure and better logistics management results in availability of the product when
farmers need it, thus ensuring demand is converted into sales. According to Parthasarathy, in
addition to land (area planted and harvested) and yield, other factors such as amount and
distribution of rainfall, cropping patterns, and size of holdings also influence fertilizer
consumption (Parthasarathy, 1994).
Forecasting in the context of fertilizers can be classified into three categories – assessment of
the potential, forecast of the demand, and forecast of the sales. While government agencies
are predominantly interested in the first two to derive their policies, companies are typically
interested in the second two. Four possible methods are identified that can be employed
towards the above stated objectives. They are “measurement of potential through need-
oriented or agronomic method”, time-series analysis, causal models, and qualitative approach.
Parthasarathy recommends using time-series methods for forecasting long-term demand and
qualitative approaches such as buyer surveys and sales force opinions to forecast short-term
demand and sales (Parthasarathy, 1994).
16
In another, more recent report, the FAO forecasted global fertilizer use for 2015 and 2030.
Agricultural inputs are tightly linked to crop production or yield and the growth of this crop
production is related to macro factors such as an increase in population and per capita income.
The positive relationship between fertilizer consumption and crop production is well
established in both developing and developed countries. The base scenario for projecting
fertilizer use assumed fertilizer use to be related to acres planted and yield. The model was a
logistic regression model, where logarithmic transformation of original independent variables
were used as inputs, since year-on-year changes in variables were used (FAO, 2000).
Finally, while Griliches’s (1958) work focused only on fertilizer usage, Tenkorang (2006)
attempted to forecast the long-term global fertilizer demand, similar to the work done by the
FAO. In addition to forecasting fertilizer consumption, which the FAO completed, Tenkorang
also attempted to estimate soil nutrient balance. Since our research will not address soil
nutrient estimation, we focus on his fertilizer estimation method. He forecasted fertilizer
demand across 182 countries split into nine regions, using data from 1962-2005. An interesting
finding is the relationship between fertilizer usages in years following a bumper harvest.
Farmers tend to have the good-year/bad-year syndrome where they feel a high yield season is
usually followed by a lower yield. Hence, they possibly lack motivation to take steps to increase
yield following a high harvest season. Accordingly, Tenkorang modeled fertilizer utilization in a
given region and time period as a liner function of current year crop output, previous year crop
output, total cultivated land, and a dummy variable to account for any structural shift in
fertilizer usage. However, since both land use and crop yield are fundamentally related, his
model suffers from multicollinearity. This was remedied by removing those independent
variables from the equation based on the variance inflation factor (VIF). The final model,
adjusted for multicollinearity, reported a high R2, with land cultivated to be most significant
variable across a majority of the regions (Tenkorang, 2006).
These studies bring out two important concepts in agricultural forecasting. First, most forecast
models used are causal models that try to relate a dependent variable such as fertilizer demand
17
to a set of exogenous variables. Second, these studies point us to some of the core factors or
independent variables that we should consider in building our model; crop yield, acres planted,
and acres harvested being the top three. However, we need to note that these models are
yearly forecasting models, whereas our interest is also in short-term forecasting.
2.3 Demand Sensing and Short-term demand forecasting
Since our project involves developing a model to help adjust short-term forecasting in response
to changing factors, we also focused on understanding the aspects of demand sensing. Demand
sensing is defined by Steutermann, Scott, and Tohamy (2012) as "the translation of demand
information, with minimal latency, to detect who is buying the product, what attributes are
selling, and what impact demand-shaping programs are having” (Steutermann, 2012). As
Robert Byrne (2012) of Terra Technology, a provider of software solutions in demand planning
and management, explains, traditional demand forecasting algorithms cannot take into account
the changes in consumer demand as they unfold, while demand sensing reacts to these
changes as they happen. Byrne also mentions the importance of data – such as economic
recessions and competitor actions – that can reflect consumer demand patterns (Byrne, 2012).
Agarwal and Holt (2005) used a short-term demand forecasting model that incorporated point
of sale (POS) data in order to fine-tune the demand forecast for a consumer goods company.
They proposed using the moving-average method, which is a time-series analysis technique in
order to forecast the next time period sales based on the POS data of the past few time
periods. While this sheds light on usage of POS data in short-term demand forecasting, we are
more interested in using causal factors in addition to POS data.
2.4 Chapter Summary
This chapter explores the terminology and challenges in big data. It also reviews the previously-
performed work on forecasting in agriculture with a focus on fertilizer. This helps us
understand the factors that should be considered and the techniques that can be applied. In
our next section, Methodology, we will explain how we approached this problem, what
variables we considered, and how we analyzed the situation.
18
3 Methodology
To explore analytical models that allow demand sensing and forecasting for the chemical
AgChem, we developed the following methodology. First, we investigated the factors that
influence the sale of AgChem. Next we broke the problem of forecasting into three sub-
problems. For each sub-problem, we identified the pertinent factors and developed an
appropriate modeling approach.
3.1 Understanding the Factors influencing Demand
In addition to conducting a literature review, we held extensive brain-storming sessions with

the supply chain and sales teams at DAS to understand the factors believed to influence the
demand for AgChem. These sessions were conducted during site visits and in telephone
conference calls.
AgChem is used in the production of corn. Thus, the agricultural factors we identified relate
specifically to corn production. The DAS sales and supply chain teams believed that factors
such as fertilizer prices, price of the product, and corn prices influence growers’ desire to invest
in AgChem. For example, an increase in the price of corn could encourage growers to invest
more in the crop by using additional yield-boosting chemicals. Similarly, retailers’ marketing
efforts were believed to influence the demand. Retailers’ willingness to market the product,
however, is influenced by their distribution infrastructure, such as liquid bulk tanks. Other
factors such as weather, including daily temperature and rainfall, were also believed to
influence demand. As discussed in the Introduction, this product is available in two varieties –
with fall being the primary sales season for V1 and spring for V2. Hence, the available
application window also has an impact on sales. Figure 4 summarizes the identified factors in a
causal loop diagram.
19
Figure 4 Factors influencing demand of the product
Some of these factors cannot be captured quantitatively. For example, while we know that
growers’ attraction to AgChem positively correlates with sales of the product, this attraction is
dimensionless. In such cases, we have to rely on the underlying factors, such as price of
fertilizer, price of the product, and price of corn, for actual analytical model development.
These factors and their presumed influence on sales are shown in Table 1. Note that corn price
and yield are lagging indicators. They determine the revenue that a farmer gets in the previous
year and hence the money he can expend on farm inputs.
20
Table 1 Factors influencing the demand and their granularity
Factor Description Granularity1
Corn Price Price of the corn. Usually measured in $ per Daily, by
bushel geography
Fertilizer Price Price in dollars per pound of fertilizer Annual
Product (AgChem) Price Sales price of the product Annual
Corn Acreage (Planted) Number of acres planted Annual, by county
Corn Acreage (Harvested) Number of acres harvested Annual, by county
Yield Bushels per acre of corn harvested Annual
Temperature Degrees F Hourly, by city
Precipitation/Rainfall Measured in in. or mm Hourly, by city
Marketing Budget $ spent in marketing per year Annual, by retailer
Distribution Installed tank capacity in gallons By retailer
Infrastructure location
# of retailers Number of retailers in a given county Annual, by county
The last three of these factors relate to DAS’s corporate strategy:

1. Marketing budget: As a part of a corporate effort to increase farmers’ awareness of the
product and the product reach, DAS collaborates with select distributors and retailers
via advertisement, promotions, and discounts. This is usually measured as the amount
of money spent in marketing efforts.
2. Distribution infrastructure: In order to increase the product availability in high growth
geographies, DAS commissions bulk tanks at retailer and/or distributor sites.
Distribution infrastructure is thus measured by the installed tank capacity.
3. Number of retailers: Since this product is generally not transported far from the retailer
location, the number of retailers in a given area could lead to increased sales via
increased product availability for the growers.
1
Granularity defines the level at which data is captured by national agencies such as USDA and hence the level at
which data is available for analysis.
21
We also tried to analyze the farmer’s decision making process when deciding whether to
purchase AgChem. The decision making process of a farmer can be broken into three steps:
1. Whether to buy AgChem: This is determined by factors such as his income from the
previous season, current corn prices, current AgChem price, production costs in the
current season, his familiarity or inclination – determined by marketing efforts and
product availability.
2. When to apply: Once a decision is made to invest in AgChem, the second choice he faces
is around the application window. A farmer can apply AgChem in the fall or spring. Even
within a season, the farmer must consider the weather since there are specific
temperature and moisture requirements for the product to be effective.
3. How much to apply: If the farmer will use AgChem and knows when, he must decide
how much of his crop will receive the product.
Once we understand the various influencing factors and this decision making process, we move
to the next step of dividing the problem and identifying model specifications.
3.2 Dividing the Problem
We decided to segment our problem into three sub-problems, or predictive models, to provide
the greatest benefit to DAS.
1. The first model will predict total annual sales on a national basis. Macro-level, yearly
variables such as average corn prices, corn acres planted, and average yield will be used
to forecast the quantity for the entire season. This will help DAS plan their annual
production schedules and understand future demand for the product.
2. The second model will determine the quantity of product that will be sold in a given
county during the entire season. We chose the county as the geographical dimension
since most of the publicly available data is measured at the county level. This will help
DAS plan the distribution of AgChem commensurate with the geographical demand.
3. The third model will be a short-term demand sensing model which will determine
weekly sales or product movement within a specified geography during the sales
22
season. Daily and weekly variables such as temperature and precipitation will be used to
develop this model.
3.3 Model Development: OLS Regression
Since our objective is to investigate the relationship between a set of independent factors and
the sales of AgChem, causal forecasting models were investigated. Accordingly, we chose the
regression method as the modeling technique.
Simple linear regression posits a linear relationship between the independent variables and the
dependent variable, whose values are to be forecasted. Equation 1 represents the simplest
form where only a single independent variable is used to predict the dependent variable.
Where Yi are the values to be forecasted and Xi, , and are the coefficients that indicate
how independent variables affect the dependent variable. , called error term or noise, stands
for the difference between the actual value of Yi and the predicted value. It is also called
ordinary least squares (OLS) regression, since the coefficients are derived by minimizing the
sum of squared errors. It is important to note that the coefficients are estimated and hence the
regression equation sometimes is written as follows, where ̂ (read y-hat) represents the
predicted value of Y and ̂ and ̂ are estimated values of and , respectively.
̂ ̂ ̂ 2
Another important concept in regression is the coefficient of determination or R2, which

measures the strength of the relationship. It represents the amount of variation in the
dependent variable that is explained by independent variables. Its value can range from 0 to 1,
with higher values usually representing better relationships (with the exception of non-linear
23
relationships). This is closely linked to the concept of residual sum of squares and total
variation. The derivation of R2 is shown in Equation 3.
SSR stands for residual sum of squares and SST stands for total sum of squares. Residual sum of
squares measures the “unexplained error” whereas total sum of squares measures the total
error as measured from the mean. Note that residuals are nothing but error terms indicated in
Equation 2.
Other important concepts of regression are t-statistics and p-values. They measure the
significance of the independent variable and usually serve as a decision criteria to include or
exclude a given variable. Typically a 95% confidence interval is used, which means a p-value of
less than 0.05 denotes significance of the independent variable.
24
4 Data and Analysis
In this chapter, we summarize the data used for the analysis. We then present the models we
developed and the insights we gained from those models. The data summary is presented in
two sections. The first section summarizes the data from our sponsor company and the publicly
available data that was used for the analysis. The second section describes the data selection
and cleaning that was applied in order to select the right data for modeling.
4.1 Data Summary
Figure 1 on page 9 of the Introduction, depicts the echelons in the supply chain that our
sponsor company is part of. As explained previously, various types of sales data were obtained
from the sponsor company. In addition to sales information, we also obtained the data listed
below.
 Inventory data for bulk tanks: Our sponsor company invested in a bulk tank
infrastructure which is monitored for inventory levels. We obtained this tank monitoring
data for use in our analysis.
 Marketing spend: DAS spends marketing dollars with retailers and distributors in order
to increase visibility of their product. This information is tracked by retailer and we
obtained this data as well.
In addition to data provided by the sponsor company, we leveraged publicly available data
predominantly from the USDA and Iowa State University. USDA sources provided us with data
on annual factors such as corn acres planted, corn acres harvested, and corn yield at a county
level. From Iowa State University’s weather data source, we downloaded data on daily high and
low temperatures, snow fall, and precipitation from September 2011 to May 2012 for various
cities. We chose the season 2011-2012 since the retailer and distributor shipment data is
available by county level only for this season.
25
4.2 Analysis and Results
The objective of our work was to use the existing public data in agriculture to identify the
relationship between macro-level and micro-level factors and the sales of the product. The
annual factors such as corn acreage and yield are classified as macro-level data whereas daily
temperatures and precipitation are classified as micro-level data. Our expectation was that
macro factors contribute to annual trends in the sales and micro factors contribute to short-
term trends in the sales.
Accordingly, the following three hypotheses are tested in this project:

1. Macro factors such as corn price, fertilizer price, corn acreage, fertilizer usage, and
corn yield influence the total annual sales of the product.
2. Geographical distribution of sales by county is linked to corn acres planted, corn
acres harvested, average yield, sales in the first few weeks of the season, marketing
dollars spent, and bulk tank capacity within that county.
3. Near-term demand (i.e. weekly or daily demand) correlates with daily weather
patterns and season-to-date sales.
To test these hypotheses, we built three main regression models. The models, along with any
additional sub-models, are described in detail in subsequent sections. In addition, we present
the model analysis and results and discuss our findings.
4.3 Preliminary Analysis
We started with a preliminary analysis of the data to understand and find patterns if any. In
addition, this analysis helped us to check the validity of our hypotheses, wherein we are
proposing correlation between various factors and the sales of the product.
We began by looking at 10 years of shipment data from DAS to its distributors to get a visual
understanding of the sales and the data characteristics. This is presented in Figure 5. We gained
three key insights from this graph. First, sales significantly increased after 2005. While we had
26
not yet weighed empirical evidence, we assumed this spike in sales was due to an increase in
corn production, driven primarily by increased ethanol production. Second, shipments seem be
fairly seasonal, with November being the peak month. This is in line with the product
application characteristics. The product is applied either after harvest in the fall or before
planting in spring. Third, there was a sizable dip in shipment volume during 2009, which was a
drought year and during a recession. This reinforces the intuition that sales of this product
could be correlated to weather and economics.
Figure 5 Increasing demand and seasonality of AgChem observed over a 10 year period
Given that we only have a few data points for each model we intend to build and that there
didn’t seem to be any unexplainable outliers, we chose not to ignore any data points.
Since we know that there are at least 3 echelons in the distribution chain, we started with a
summary plot of shipments from DAS to distributors, distributors to retailers, and retailer to
growers for the calendar year 2011. In accordance with intuition, Figure 6 shows that shipments
from manufacturer to distributors happen before the other two echelons, thus re-assuring us of
the data sanctity.
27
Figure 6 Product flow between echelons
4.4 Model-1: Annual Sales Model
The first model, which we call “annual model,” posits a relationship between annual sales of
AgChem and these independent variables: average corn price received, corn acres planted,
average yield, fertilizer price, and fertilizer usage. The plot of these annual variables is shown in
Figure 7. As the plot shows, corn price, fertilizer price and price of AgChem appear to be
positively correlated with annual sales of AgChem.
28
Figure 7 Annuals sales trending along with fertilizer and corn price
Next, we computed correlation coefficients for these variables using SAS’ statistical software,
JMP. The results are presented in Table 2 below. The correlation coefficient matrix shows that
the sales are strongly correlated with corn price, fertilizer price, AgChem price, and corn acres.
Out of these four independent variables, AgChem price is available from 2006 only. Hence, this
variable was excluded from the regression model and only the remaining three variables were
used. The results of regression developed using SAS JMP are summarized in Table 3.
29
Table 2 Correlation coefficient matrix for all variables
AgChem Fertilizer Corn Corn Yield Fertilizer AgChem

Sales Price Price Acres Usage Price
AgChem Sales 1
Fertilizer Price 0.83 1
Corn Price 0.99 0.82 1
Corn Acres 0.89 0.77 0.90 1
Yield -0.13 0.27 -0.10 0.01 1
Fertilizer Usage 0.06 -0.02 0.005 0.10 -0.15 1
AgChem Price 0.72 0.67 0.73 0.52 -0.28 0.29 1
Table 3 Regression results of the annual model
Estimate Std. Error t-ratio P-value

Intercept -874773.65 1961678 -0.45 0.6662
Fertilizer Price ($ per lb.) 443.83 653.80 0.68 0.5143
Corn Price (dollars) 872405.76 111910.7 7.8 <.0001
Corn Acres -0.012 0.027 -0.46 0.6554
Fit Statistics
Adjusted R2 0.97
Mean Squared Error 258868
The first column lists the variables and the second column gives the value of the coefficient. We
are interested in the last column, p-value, as it indicates the significance of a given variable in
the regression relation. A p-value of less than 0.05 is considered to be significant. As expected,
corn price turns out to be a very significant factor. But surprisingly, fertilizer price and corn
acres have a p-value much greater than 0.05, indicating that they might not be significant. The
variable corn acres also has a negative sign for coefficient. But given, that the p-value of this
coefficient is greater than 0.05 we suspect that corn acres might be correlated to other
independent variables, i.e., the model has more explanatory variables than needed.
30
Subsequently, we dropped the two terms that have insignificant p-values and developed the
regression model with just corn price as the independent variable. The results are shown in
Table 4. As can be expected, both the intercept and corn price turn out to have significant p-
value and there is virtually no change in adjusted R2 value.
Table 4 Regression results of the annual model with only significant factors

Intercept -1713408.67 155083.85 -11.05 <.0001
Corn Price (dollars) 872628.97 39583.75 22.05 <.0001
Fit Statistics
Adjusted R2 0.97
This fits with our intuition and the graphical analysis we did earlier. One way to explain this
relationship is to say that as corn prices go up, farmers have an incentive to produce and sell
more as well as additional capital to spend on inputs. Hence, a farmer would be interested in
spending more in his production and will be willing to invest in AgChem, which can boost yield.
However, it is difficult to believe that corn price is the only driver of sales and that corn acres
and fertilizer price have no role to play. As we can observe in Figure 7, sales are positively
correlated with the price of fertilizer. One possible explanation is that as price of fertilizer
increases the grower wants to insure his expenditure and revenues and hence will buy AgChem.
To evaluate whether increased corn acres planted leads to increase in sales we need to know
the market penetration of the product. A higher penetration means farmers are aware of the
product and hence an increase in acreage might lead to an increase in sales. While we cannot
fully explain the insignificance of fertilizer price and corn acres, we need to notice that this
regression was developed with just 13 data points and, therefore, might not accurately depict
the relationship between dependent and independent variables. Further analysis with more
data points would validate this result.
31
4.5 Model-2: Geographical Sales Model
Next we developed a model to explain the geographic spread of the sales. Our hypothesis is
that sales in a given geographical location are correlated to acres harvested in the area, yield in
the area, number of retailers in the area, and quantity sold in the first two months of the
season. It is important to understand the rationale for our hypothesis.
Usually the product is applied either after harvest in the fall or before planting in the spring, for
the next growing season. For example, product is applied in September 2011 – April 2012 for
the corn season that starts in April 2012 and ends in November 2012. Our intuition is that if a
farmer has a good yield and lots of acres planted in the current season (April 2011 – November
2011), then his income would be higher, as would his ability to spend more on farm inputs.
Hence, income from this season will affect his decision to purchase the product or not.
Similarly, more retailers in a given area will produce higher marketing power and accessibility of
the product, thus leading to higher sales. Finally, our interviews with our sponsor team
indicated that typically higher sales in the first two months indicate a higher season overall.
Hence sales in the first two months is included as an independent variable. However, we
understand that adding the first two months sales as explanatory variable will automatically
lead to higher R2 due to very high correlation between total season sales and the first two
months’ sales. Thus, we first evaluate the model without this variable and then add it. The
model regresses the sales in September 2011- April 2012 as a function of these independent
variables.
We considered a number of other independent variables and tested for their strength using p-
values in ordinary least squares regression. We dropped the variables whose p-values were
insignificant, leaving only the four variables discussed above in the model. Each one of these
was tested individually in a regression model for its strength before the insignificant variables
were dropped. The complete list of all variables tested and their status (included or not
included) is summarized in Table 5.
32
Table 5 Summary of factors for geographic model
Variables Granularity W/ Dist to retailer data

Distributor sales by geography By county, annual Dependent (Y)
(i.e. to retailer)
Corn acres planted (2011) By county, annual Not Significant - Excluded
Corn acres harvested (2011) By county, annual Significant
Yield (2011) By county, annual Not Significant - Excluded
Increase in yield (Y-o-Y) By county, annual Not Significant - Excluded
# of retail locations By county (aggregated ) Significant
Bulk tanks By county (aggregated ) Not Significant - Excluded
Annual rainfall (2011) By county, annual (crop season) Not Significant - Excluded
Historical sales By county Not Significant - Excluded
Sales in Sep-Oct 2011 By county Significant
Quantity purchased by grower By county (aggregated ) Not Significant - Excluded
Farm size (e.g. customer size) By county (aggregated ) Not Significant - Excluded
We made two key decisions while building this model. The first one was which sales data to
use. We had two sets of sales data: sales from distributor to retailer and the retailer to farmer.
The second decision was the geographic level for data aggregation (e.g. city vs. county vs.
state). We choose to use distributor to retailer sales since it was more complete than the
retailer to farmer data and it also had the retailer ship to location. Retailer location is a fairly
good approximation of actual grower application location, since growers typically don’t travel
very far to obtain the product. We then aggregated the data to the county level. Given that we
didn’t have the exact city/location of application (i.e. grower application city), we felt
aggregating by county was the best choice based on our discussions with the sponsor company.
The results of the regression on the statistical analysis package SAS JMP are displayed in Table
6. In total, we have 134 counties, hence 134 data points. We can see that acres harvested and
number of retailers are significant.
33
Table 6 Regression results of geographic model

Intercept 74.08 1618.46 0.05 0.96
Harvested 0.03 0.01 2.92 0.0041
Count of Retailer 311.63 48.36 6.44 <.0001
Fit Statistics
Adjusted R2 0.38
Mean Squared Error 8149.8
Next, we included the first two months’ sales as an explanatory variable. The results are
reported in Table 7. We can see that in addition to acres harvest in prior season and number of
retailers, the sales in first two months also turns out to be significant as can be expected. While
we expected certain additional factors, specifically bulk tank capacity and annual rainfall, to be
significant, they are not. Even though yield and intercept term have a p-value that is higher
than 0.05, they are left in the model for sake of completeness.
Table 7 Regression results of geographic model with season start sales included

Intercept 7187.19 3936.08 1.83 0.0701
Harvested (2011) i.e. prior season 0.026 0.010 2.41 0.0176
Yield (BU/Acre) (2011) i.e. prior season -51.02 28.40 -1.8 0.0748
Count Of Retailer 235.56 42.18 5.58 <.0001
Quantity sold in September & October (2011) 1.72 0.23 7.34 <.0001
Fit Statistics
Adjusted R2 0.55
34
While the R2 is not very high, the model makes intuitive sense. A higher harvest will leave the
farmer with more money which can drive sales. Similarly, as number of retailers increases the
product reach increases and we would expect sales to increase. Also note that we didn’t discard
any data. A careful search for outliers will help us get a better R2. Finally since we didn’t have
exact location of product application, we used retailer location and then aggregated it to
respective counties. This approximation might also have skewed the result we presented
above. Nonetheless, the above model presents a significant finding that harvest, number of
retailers, and sales in the first few weeks act as strong indicators of total season sales.
4.6 Model-3: Short-term Demand Model
Next, we developed a short-term demand model that would correlate weekly sales with
weather related variables. The decision of a farmer about when to apply the product depends a
lot on weather factors such as temperature and precipitation. If the weather is too cold and
hence is not conducive for the farmer to go into the field, then he will not apply the product.
Also, for application in the fall the temperature has to be no more than 50⁰ F and trending
down. In light of these details, we conducted regression analyses on weekly sales with average
weekly temperature and precipitation. Since it was not possible to do this analysis for all the
1000+ cities in the data, we chose to do this only for two of the top selling cities. Before we
present our detailed analyses on the weather relationships, we present a few key insights we
gained while analyzing the data.
4.6.1 Timing of Product Application
As described in the Introduction, there are two variants of the products that farmers can
choose to apply. The first variant AgChem V1 is applied during the fall (after harvest) or in the
spring (before planting). The second variant, AgChem V2 is applied only in the spring. For this
analysis we considered only the first variant, AgChem V1, since this variant accounts for more
than 80% of the sales and is also the oldest.
In our first analysis we focused on identifying whether there was a pattern to when farmers
chose to apply the product. We have 5 years of distributor sales data that gave us the retailer
35
cities where the product was shipped. Using this, we plotted the quantity sold in fall and spring
as a pie chart across cities over the years. The results indicate a trend that some geographies
are predominantly fall application communities whereas the rest tend to be spring application
communities. In Figure 8, green indicates quantity sold during fall season (September –
December) and red indicates quantity sold in spring season (January – April). We see that most
of Iowa and northern Illinois are predominantly green whereas southern Illinois and Indiana
tend to be red indicating these are spring-application states. We plotted data for subsequent
years as well (refer to appendix-1) and this pattern seems to hold true across years.
Figure 8 Geographic plot of fall and spring quantities shipped (2007)
Next, we analyzed whether we could identify any events that could potentially act as leading
indicators of the sales.
36
4.6.2 Harvest date as a trigger point for fall application
The first insight we gained is how harvest date acts as a trigger point for application in fall. For
the product to be applied, the corn crop has to be harvested. This leaves the farmer with a
choice of applying it in fall or in spring. We first found the peak week of sales for each city. This
will be the week where maximum sales happened. We then plotted the quantity sold in the
peak week of each city from September 2011 to April 2012. This is shown in the Figure 9. Week
48 (red line) represents the week when 100% of the corn was harvested. As we can see, there is
very high demand (sales) from week 49 to week 50 and significant demand around week 13.
Note that week 41 through 53 represent the fall of 2011 whereas week 1 through 17 represent
the spring of 2012. This means a lot of cities sell the product in the first couple of weeks
following harvest. Additionally, there are a considerable number of cities where peak sales
happen in the spring. Also, the total peak week sales account for 31% of the total season sales.
This means that essentially 1/3 of all sales happen during the peak week of each city. The
manufacturer can benefit immensely by forecasting the peak week and positioning inventory to
serve these spikes in demand. And harvest date could act as a trigger point for sales, essentially
helping us forecasting peak weeks.
Figure 9 Peak sales occur within 1-2 weeks after harvest completion
37
Next, we developed a regression model to analyze how daily temperature and precipitation
influence sales.
4.6.3 Temperature and precipitation as predictors of weekly sales
We have point of sale data from the retailers for the 2011-2012 season. We used this data to
obtain aggregate sales by city, for each week. This data contains nearly 1000 cities. Since we
could not analyze all these cities we decided to pick one city with the highest sales in fall and a
second one with the highest sales in spring. Thus we end up with Ames, Iowa and Avon,
Indiana.
4.6.3.1 Results for Ames, IA
Note that Ames is highest in the fall. The model uses weekly sales as the dependent variable
with average weekly temperature and average weekly precipitation acting as independent
variables. The regression results for Ames, Iowa are shown in Tables 8 and 9.
38
Table 8 Weekly quantity sold data for Ames, IA
Week Start Date Quantity sold (quarts) Average Temp (F) Average Precipitation (mm)
9/25/2011 147.5 62.3 0.0
10/2/2011 75.0 65.0 0.0
10/9/2011 165.0 59.5 8.5
10/16/2011 126.9 44.0 0.0
10/23/2011 1124.7 51.7 0.0
10/30/2011 2845.4 45.4 3.6
11/6/2011 6007.0 41.7 7.5
11/13/2011 5397.6 40.4 0.0
11/20/2011 7687.4 41.8 0.0
11/27/2011 11611.7 31.6 0.0
12/4/2011 11712.7 19.4 0.1
12/11/2011 7753.3 37.0 4.7
12/18/2011 3151.0 31.3 0.1
12/25/2011 2371.6 36.6 1.3
1/1/2012 25.6 45.5 0.0
1/8/2012 204.1 31.3 0.3
1/15/2012 935.3 8.8 0.7
1/22/2012 70.9 59.5 0.0
1/29/2012 34.5 70.0 0.0
2/5/2012 337.1 58.8 0.0
2/12/2012 1290.4 59.0 1.1
2/19/2012 1882.6 47.4 2.8
2/26/2012 855.4 53.1 3.9
3/4/2012 629.9 60.1 2.8
39
Table 9 Regression results of weekly model using entire season data

Intercept 8755.97 2140.20 4.09 0.000522
Average Temp -134.86 44.42 -3.03 0.006279
Average Precipitation 128.43 269.24 0.48 0.63828
Fit Statistics
Adjusted R2 0.24
80 14000
70 12000
Quantity sold in Gallons

60
10000
Temperature (F)
50
8000
40
6000
30
4000
20
10 2000
0 0
Week Start Average Temperature
Average Precipitation
Qunatity sold
Figure 10 Weekly sales correlate with temperature for Ames, IA
40
Figure 10 indicates that significant reductions in air temperatures coincide with jumps in sales.
Each time there is a major drop in temperature, there is a large peak in sales. We cannot,
however, conclude that the colder it is the more sales will occur. If this was true, sales would
be highest during the coldest part of the year. We can see in the figure that this is not the case.
The peaks in sales are not proportional to the drops in temperature.
We see that the adjusted R2 is very low and precipitation is not at all significant, as its p-value is
much greater than 0.05. Since we know that Ames is predominantly a fall application city, we
reran the model by including fall weeks only. The results in Table 10 indicate much better R2
value; however precipitation still is not a significant factor.
Table 10 Regression results of weekly model using only fall data

Intercept 16186.97 2663.85 6.08 0.0001
Average Temp -270.46 59.19 -4.57 0.0010
Average Precipitation 73.03 247.74 0.30 0.7741
Fit Statistics
Adjusted R2 0.61
4.6.3.2 Results for Avon, IN
Next we ran a similar model for Avon, IN. Figure 11 shows that Avon is a city with high spring
sales. It is also one of the top-30 cities for overall sales. The data and results of the regression
are shown in Tables 11 and 12.
41
Table 11 Weekly quantity sold data for Avon, IN
Week Start Date Quantity sold (quarts) Average Temp (F) Average Precipitation (mm)
9/25/2011 2 53.5 0
10/2/2011 130 59.5 0
10/9/2011 28 62 1.4
10/16/2011 15 45 21.6
10/30/2011 20 45.5 0.5
11/6/2011 41 43.25 2.3
11/13/2011 1225 46.4 2.74
11/20/2011 680 44 0.4
11/27/2011 3561 36.4 8.88
12/4/2011 3169 33 3.04
12/11/2011 985 38 5.075
12/18/2011 541 45.25 9.4
12/25/2011 1230 39.17 0.43
1/8/2012 31 23.25 0.25
1/15/2012 123 29.5 2.05
1/22/2012 8 43.5 17.8
2/12/2012 68 29.75 0.65
2/19/2012 249 37.5 0
2/26/2012 282 41 0
3/4/2012 579 42.2 0.06
3/11/2012 2587 64.1 1.02
3/18/2012 10454 67.29 2.47
3/25/2012 16134 54.75 1.95
4/1/2012 7410 55.17 0.55
4/8/2012 11354 48.42 2.62
4/22/2012 13163 52.33 1.13
42
Table 12 Regression results of weekly model using entire season data

Intercept -5495.09 3947.28 -1.39 0.1766
Average Temp 196.40 82.24 2.39 0.0251
Average Precipitation -89.53 169.02 -0.52 0.6011
Adjusted R2 0.14
Figure 11 Peak sales occur in spring for Avon, IN
Similar to the case for Ames, IA, Figure 11 indicates that lower temperature in fall corresponds
to higher sales. However, Avon is a city where peak sales happen during spring. Hence, we
reran the regression model with data from only spring weeks. The results are presented in
Table 13.
43
Table 13 Regression results of weekly model using only spring data

Intercept -8560.1 4940.15 -1.73 0.1110
Average Temp 311.15 103.08 3.018 0.0116
Average Precipitation -180.88 297.80 -0.60 0.5559
Adjusted R2 0.36
The results indicate that temperature is a significant factor but precipitation is not. While there
is an improvement in the value of R2, it is not strong enough to be conclusive in Avon, IN.
We discussed these results with our sponsor company and concluded that temperature acts as
trigger point similar to harvest dates but temperature by itself not a driver for sales. For
example, in fall a temperature of 50⁰ F is desirable and once this threshold is reached the
application can start. The application can continue until the ground freezes. However, a
decreasing temperature doesn’t necessarily translate to increase sales through the winter. In
summary, temperature acts as an on-off switch that triggers sales rather than being a
continuous variable affecting sales.
44
5 Conclusion
5.1 Models and Significant Variables
Our project scope was to identify how DAS can use the vast publically-available data
surrounding agriculture to improve their supply chain processes and decisions. We set out to
do this through the development of predictive models to forecast demand of AgChem on
various time and geographical horizons. We sought to identify factors which could help us
predict the total annual demand, the annual demand by geographical sub-sets (county), and a
short-term demand sensing model to predict when this demand will occur. We expected that
there was enough agricultural data available for us to find some related factors.
For the total annual demand, we found two factors which appear to be significantly correlated;
the average price of corn received by growers the previous harvest and the price of fertilizer.
We used annual AgChem sales data from DAS for the years 1999 through 2012.
For the geographical demand, we identified that the number of retailers in the area, the corn
acres harvested in that area, and the first two months of AgChem sales are correlated to the
amount of AgChem which is sold throughout the season. We also identified an interesting
pattern about the sales during the spring versus the fall. We found that through the past five
years, sales have been heavier in the spring in southern Illinois, through Indiana, and into Ohio.
However, in the rest of the major corn-producing regions, AgChem is sold predominately in the
fall.
For the short-term demand sensing, there are two triggers for the beginning of the sales period
for fall application. The previous crop needs to be harvested and out of the field, and the
temperature must be below 50⁰ F. For the year we analyzed, the temperature was not an issue
as it was already cold enough when the crops were harvested. For that year, we discovered
that the largest volume of fall sales at the retailer locations occurs generally within the two
weeks following the corn harvest. This short window following the harvest accounted for
roughly 40% of the fall sales. See Table 14 for these variables.
45
Table 14 Models and variables
Model Dependent Variable Independent Variables - Significant
Model-1: Annual sales nation-wide Corn price, Fertilizer price
Annual Model (gallons/year)
Model-2: Annual sales by county Acres harvested, Number of retailers,
Geographic Model (gallons/county) Sales in the first two months
Model-3: Sales by week in a given city Average weekly temperature, Harvest
Weekly Model (gallons/week) completion date
5.2 Implied Benefits to DAS
This research and its results can benefit DAS in a variety of ways. DAS can better understand
the ordering trends of each link in the distribution chain. They can quantifiably see when
orders and shipments happen from them to the distributors, from the distributors to the
retailers, and from the retailers to the growers. With this information, DAS can make decisions
about the production and distribution of AgChem.
With a better understanding of the factors driving annual demand for the product, DAS can
make production plans which balance the best utilization of their manufacturing assets and the
expected profits from the sales of AgChem.
DAS can estimate where demand might be higher and lower in a given year based on the
identified factors. They can also gain additional insight into how the sales period will progress
within those specific geographies. Knowing this, they can make more strategic logistics
decisions to lower their distribution costs and utilize their inventory more effectively.
5.3 Next Steps and Additional Research
We have identified a few additional steps which could add to the research we have done, which
we did not have time to complete. We feel that these additional steps could add clarity to
46
some of our results and provide more conclusions which will benefit our understanding of
AgChem demand.
As noted above, we identified some geographic trends around whether AgChem is applied in
the spring or fall. We feel that additional research in this area could uncover factors driving this
trend. This research could seek to explain whether climatic factors such as moisture and
temperature are affecting this trend. Other considerations might be whether the size of a
retailer has any effect on the season in which they realize the most sales.
For our short-term demand sensing model, we only looked at one year of retailer sales. It
would be valuable to analyze additional years as well. This might include seeing if the weather
conditions ever hold back the application of AgChem even after the crops have been harvested.
It would also be useful to determine whether the timing of harvest affects the volume of sales
realized in the fall in comparison to those in the spring, assuming that the temperature has
reached 50⁰ F.
5.4 Considerations for Future Applications
In addition to the afore-mentioned conclusions, we learned some valuable lessons which DAS
should consider if they choose to use this type of analysis in the future. These considerations
include identifying reliable, consistent data sources, collecting data for longer periods of time,
and possibly applying the analysis to more granular geographies.
5.4.1 Data Quality
We received lots of data from DAS to use in these analyses. This data included DAS sales and
shipments to distributors, distributor sales to retailers, and retailer EDI sales to growers. DAS
also provided grower demographic data, retailer locations, AgChem pricing, and other
applicable data. We also searched through masses of publically-available data. All of this data
was valuable and useful in conducting this research. The format, granularity, and
comprehensiveness of that data, however, could be improved to streamline this process in the
47
future. We felt that excessive time was required to clean the data and that in the current
formats it would be inefficient and unrealistic for DAS employees to perform the analysis.
If DAS wishes to carry this analysis forward, we recommend that they continue to work with
distributors and retailers to obtain sales data which is as complete as possible. If they receive
more complete data, they will be able to track exact volumes throughout the entire distribution
chain. This will improve the accuracy of any analyzes they perform.
We also recommend that DAS determine the granularity at which they will analyze the data.
They should consider time (daily, weekly, etc.) and geography (city, county, state). They should
then identify their internal data sources and request that the data be formatted in a standard
manner. While making these requests and formatting the data, DAS should consider how to
incorporate external data including commodity prices and weather. Matching the data in a
standard format will allow for much more efficient analysis with less risk of calculation error.
For more accurate analyses, DAS could also likely find better sources for weather readings and
other useful, external data. The publicly-available sources we used in our project were
sufficient for our analysis. However, we had to approximate some weather data for smaller,
more rural areas. If DAS would like to streamline the data gathering process while increasing
accuracy, they should be able to obtain more accurate data in a more convenient format.
5.4.2 Application City not Retailer City
The most granular sales data we received was the retailers’ EDI sales data to the growers. This
data was helpful in establishing the last point of sale in the AgChem distribution chain.
However, we had to make assumptions about the growers’ locations based on the retailer
location. We assumed that growers would not transport AgChem too far from the retailer.
Thus, we used the retailer location as an approximation for the actual application site.
This level of granularity is not a problem if we are only trying to forecast sales through the
retailer level using historic sales, commodity prices, etc. But if we are trying to use climatic
factors and grower behavior to forecast specific geographic demand, we would need one of
two things:
48
1. We would need to know how far growers were actually transporting AgChem from the
retailer locations. If it is only a short distance, we can continue with our current
assumption and use the retailer location for weather data etc. If we identify that
growers transport the product a great distance from the retailer, the weather data for
the retailer city would not be an appropriate factor for indicating short-term demand.
2. If the growers are transporting the product far enough away from the retailer that the
weather patterns could be different, then we would need further information about the
actual application site.
5.4.3 EDI for Multiple Years
DAS provided valuable retailer EDI sales data from September 2011 to April 2012. With this
data, we were able to identify specific trends and correlations within that sales period. Since
the sales of agricultural products varies from year to year due to economics, weather, etc. it
would be useful to have additional years of retailer sales data. This data could be used to verify
that the relationships we identified in the provided data hold true through other years.
49
Appendix
The following plots show the pattern described in section 4.6.1. Southern Illinois and Indiana
seems to be predominantly spring application states where as Iowa and Northern Illinois are fall
application states.
Figure 12 Geographical plot of fall and spring quantities shipped (2008-2009)
50
51
52
References
About Us | Dow AgroSciences. (n.d.).Dow AgroSciences - A Dow Chemical Company. Retrieved

March 3, 2013, from http://www.dowagro.com/about/
Agarwal, A. & Holt, G. (2005). Reducing Inventory by Simplifying Forecasting and

Using Point of Sale Data. MIT M. Eng. thesis
Brynjolfsson, E., Hitt, L. M. & Kim, H. H. (2011). Strength in Numbers: How Does Data-Driven
Decisionmaking Affect Firm Performance?. Retrieved February 10, 2013 from
http://ssrn.com/abstract=1819486
Byrne, Robert. "Beyond Traditional Time-Series: Using Demand Sensing to Improve Forecasts in
Volatile Times." Journal of Business Forecasting. Summer (2012): 13-19. Print.
FAO. (2000). Fertilizer requirements in 2015 and 2030. Food and Agriculture Organization of the
United Nations. Rome, Italy
Griliches, Z. (1958). The Demand for Fertilizer: An Economic Interpretation of a Technical

Change. Journal of Farm Economics, 40(3), 591-606.
McAfee, A., & Brynjolfsson, E. (2012). Big Data: The Management Revolution. Harvard Business
Review, October, 61-68.
Mehring A., & Shaw, B. (1944). Relationships Between Farm Income and Farmers' Expenditures
for Fertilizer and a Forecast of the Commercial Demand for Fertilizer in 1944 and 1945, by
States. American Fertilizer
National Research Council. (1995). Research at Land Grant Colleges of Agriculture. Colleges of
agriculture at the land grant universities a profile (p. 58). Washington D.C.: National Academy
Press.
Parthasarathy, N. S. (1994). Demand forecasting for fertilizer marketing. Food and Agriculture
Organization of the United Nations.
Products & Solutions | Dow AgroSciences. (n.d.). Dow AgroSciences - A Dow Chemical Company.
Retrieved March 3, 2013, from http://www.dowagro.com/products/
Steutermann, S., Scott, F., & Tohamy, N. (2012). Building an Effective Demand-Planning
Process. Gartner, Web. Retrieved April 2, 2013, from http://www.gartner.com/id=2085615
53
Tenkorang, F. (2006). Projecting World Fertilizer Demand in 2015 and 2030. Ann Arbor:
ProQuest
Vance, A. (2012, March 22). Climate Corp. Updates Crop Insurance via High Tech. Business
Week, Web. Retrieved February 10, 2013, from http://www.businessweek.com/articles/2012-
03-22/climate-corp-dot-updates-crop-insurance-via-high-tech
Vail, E. (1927). Prices of Fertilizer Materials and Factors Affecting the Fertilizer Tonnage. Cornell
University Ph.D. thesis. Ithaca
54

Satya Dhavala Derek Smith Thesis

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

Satya Dhavala Derek Smith Thesis

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Satya Dhavala Derek Smith Thesis

Uploaded by

Copyright:

Available Formats

Using Big Data for Decisions in Agricultural Supply Chain

B.Tech. Electronics and Communications

Submitted to the Engineering Systems Division

Master of Engineering in Logistics

Massachusetts Institute of Technology

Submitted to the Engineering Systems Division

Thesis Supervisor: Dr. Bruce C. Arntzen

1.1 Agricultural Data

1.1.1 Applications of this Data

1.2 Agricultural Supply Chain

1.2.1 Product Flow through the Agricultural Chemical Supply Chain

Manufacturer Distributor Retailer Grower

Figure 1 Agricultural chemical product flow

Manufacturer Distributor Retailer Grower

Inventory Levels POS Data Demographics

Figure 2 Agricultural chemical information flow

1.2.3 Our Sponsor Company

1.3 Problem Statement

1.3.1 AgChem – The Selected Product

AgChem is an agricultural chemical and is manufactured in two formulations, each under a

1.4 Thesis Outline

2.1 What is Big Data?

2.2 Forecasting in Agriculture: Review of Fertilizer Forecasting Studies

2.3 Demand Sensing and Short-term demand forecasting

2.4 Chapter Summary

3.1 Understanding the Factors influencing Demand

In addition to conducting a literature review, we held extensive brain-storming sessions with

The last three of these factors relate to DAS’s corporate strategy:

3.2 Dividing the Problem

3.3 Model Development: OLS Regression

Another important concept in regression is the coefficient of determination or R2, which

4.1 Data Summary

Accordingly, the following three hypotheses are tested in this project:

4.3 Preliminary Analysis

4.4 Model-1: Annual Sales Model

AgChem Fertilizer Corn Corn Yield Fertilizer AgChem

Table 3 Regression results of the annual model

Estimate Std. Error t-ratio P-value

Estimate Std. Error t-ratio P-value

Variables Granularity W/ Dist to retailer data

Estimate Std. Error t-ratio P-value

Estimate Std. Error t-ratio P-value

4.6 Model-3: Short-term Demand Model

4.6.1 Timing of Product Application

Figure 8 Geographic plot of fall and spring quantities shipped (2007)

4.6.3 Temperature and precipitation as predictors of weekly sales

4.6.3.1 Results for Ames, IA

Estimate Std. Error t-ratio P-value

Quantity sold in Gallons

Week Start Average Temperature

Figure 10 Weekly sales correlate with temperature for Ames, IA

Table 10 Regression results of weekly model using only fall data

Estimate Std. Error t-ratio P-value

4.6.3.2 Results for Avon, IN

Estimate Std. Error t-ratio P-value

Figure 11 Peak sales occur in spring for Avon, IN

Estimate Std. Error t-ratio P-value

5.2 Implied Benefits to DAS

5.3 Next Steps and Additional Research