國立臺灣大學管理學院國際企業學研究所 碩士論文
Graduate Institute of International Business College of Management
National Taiwan University Master Thesis
以層級貝氏模型預測廠商異質性下之銷售量
—以晶片廠商為例
Forecasting Sales Volume of Industrial Product with Firm Heterogeneity—Case of Mobile Phone Chipset
林平 Ping Lin
指導教授：任立中 博士 Advisor: Lichung Jen, Ph.D.
中華民國 99 年 6 月
Abstract
Sales forecast has been an integral part of business planning, especially to
the high-tech industry where product life cycle is short and intensive capital
expenditure is required. However, forecasts in industrial product are usually less
accurate and companies tend to adopt different forecast practices compared with
consumer product industry. Coupled with the fact that the market structure for
high-tech industry has been undergone several waves of evolutions, forecast method
should be modified to adjust for improvements.
This research paper proposed using a 2 level Hierarchical Bayesian Model
that takes customer heterogeneity into account. The first level will address the
aggregate factors affecting sales in the industry, whereas the second level utilizes firm
specific factors to explain variations among customer purchasing behavior. Empirical
analysis was accomplished with the data that recorded sales volume and firm
attributes of 8 key accounts from an IC design company.
Key word: forecast, industrial product, Hierarchical Bayesian
Contents
Abstract ... i
Chapter 1 Preface ... 1
1.1 Background ... 1
1.2 Research Purpose ... 3
1.3 Framework ... 4
Chapter 2 Literature Review ... 5
2.1 High-tech Industry Overview ... 6
2.2 Industrial Product Forecast ... 12
2.3 Heterogeneity Among Purchasers ... 15
2.4 Multi-level Forecast Model ... 18
Chapter 3 Research Method ... 20
3.1 Hierarchical Bayesian Method ... 20
3.2 Markov Chain Monte-Carlo Methodology ... 23
Chapter 4 Empirical Analysis ... 24
4.1 Sample Summary ... 24
4.2 Model Construction ... 29
4.3 Results & Forecast ... 37
Chapter 5 Conclusion ... 50
5.1 Managerial Implications & Suggestions ... 50
5.2 Limitation & Future Directions ... 51
References ... 53
Exhibits
Exhibit 1 Account Information ... 26
Exhibit 2 Descriptive Statistics of 8 Accounts (sales quantity in thousands) ... 27
Exhibit 3 Initial Macro-economic Variable Pool ... 31
Exhibit 4 Aggregate Variables ... 32
Exhibit 5 Correlation Matrix of Aggregate Variable and Sales ... 33
Exhibit 6 Individual Variables... 35
Exhibit 7 Summary Statistics for X ... 38
Exhibit 8 Summary Statistics for Z ... 39
Exhibit 9 Attribute Information ... 40
Exhibit 10 Estimates of ... 44
Exhibit 11 Estimates of ... 45
Exhibit 12 Posterior Estimate of ... 46
Exhibit 13 MLE Estimates of ... 48
Exhibit 14 Fitted Statistics ... 48
Graphs Graph 1 Monthly Sales Quantity (in thousands)... 28
Graph 2 Forecast Result (8 accounts) ... 49
Chapter 1 Preface
1.1 Background
Sales forecast has long been considered an integral part of business
planning, since it is one of the most important inputs for managerial decision making
process and affects several aspects of organization strategy. In the short-run, it
influences supply side behavior such as manufacturing and pricing; whereas in the
long run, firms must choose suitable R&D and capital expenditure based on their
strategic goals under its forecasted growth potential and demand. That is to say, the
accuracy of sales forecast has a drastic impact on successful business operation that
maximizes profit and minimizes cost or investment loss. Makridakis (1990)
discovered that among the 175 companies he surveyed, 92 percent of them deemed
forecasting is a major attribute for their company’s success.
This is even more so for high-tech sector. This industry has a notably
shorter product life cycle due to the well-established Moore’s Law—in 1965, which
states that the number of transistors that can be placed on an integrated circuit will
double every 18 months and thus prompting dramatic increase of computing power
and innovation for IC related products. Under this extreme pace of product
replacement, near-term excess inventory as a result of imprecise sales forecast
translates to loss from product obsolescence. In the long run, false expectation to
future growth in demand will lead to untimely investment that drastically raises the
degree of operational risk and puts a heavy burden on the company financially. As
such, management in high-tech sector eagerly seeks more effective forecast methods
and variables to capture the change in demand.
Important as it is, sales forecast is never an easy task judging on the ever
changing business environment and fast evolving market structure. To say the least,
the boundary between each market has been broken and a lot of international players
are now joining the industry and competing with other international suppliers for
buyers form different regions of the world, making macro-economic performance and
regulation in foreign countries relevant issues to consider when forecasting company
sales.
To make things even more complicated, the structure within a certain
market has becoming more and more complicated as a lot of the traditional business
models gradually lose its profitability and new ways of formulating a value chain
emerge, bringing more market players into the industry. As a result, the interplay
between supply and demand can seldom be accurately captured with a general model
ignoring the different behaviors of each market player.
These factors—jumps in technology and product, cross-national operation,
and different forms of purchaser-supplier relationship—have rendered forecasting a
challenging but crucial task that this research attempt to probe into. Both related
theories and empirical studies will be touched on to present a comprehensive view.
1.2 Research Purpose
The purpose of this study can be viewed as an endeavor in answering these
propositions:
A. Reviewing existing forecasting theories to draw useful insights into
current forecasting
B. Understanding features of the high-tech industry and how current
forecasting model should address these business changes
C. Discovering the external drivers affecting general sales volume in
the industry
D. Discovering the respective attributes that influence purchasing
behavior across different accounts in a certain company
E. Proposing a forecasting methodology that exploits these relevant
information in the most accurate way
1.3 Framework
The research is composed of 5 chapters. Following this chapter that
provides basic introduction of the research motive, chapter 2 will outline the academic
basis of this research covering fields such industry overview, forecast theories and
methods, and would then focus specifically on the forecast of industrial product as the
primary concern. Chapter 3 discusses the research methods utilize in this paper,
including two statistic parameter estimation methods—Maximum Likelihood
Estimation Method(MLE method)and Hierarchical Bayesian Method (HB method)—
and one algorithm, Markov Chain Monte-Carlo method. Empirical analysis will be
recorded in Chapter 4 along with summary reports of the data, construction of the
forecasting model, and forecasting results. Finally, Chapter 6 concludes with
managerial implications and some suggestions on modifying existing forecasting
practice. Limitation on the research and future directions for further study will also be
covered in this chapter.
Chapter 2 Literature Review
This chapter will first look into the specific industry sector of
interest—namely the high-tech industry. By examining the operation environment,
market structure, and customer characteristics, the key drivers and suppressants can
be identified for input screening for model building in the next chapter.
Then, a review on the development of in-corporate forecasting will serve as
the start point for constructing a model in the aim of providing companies a better
capture of their sales outlook. Finally, a comparison between different forecasting
forms will be conducted to augment the discussion and solidify the choice of the
model in this paper.
2.1 High-tech Industry Overview
High tech industry includes a wide range of production business and there
are no strict criteria to separate high-tech business to those of “non-high-techs.”
Nevertheless, OECD methodology in classifying product category uses R&D
intensity— industry R&D expenditure divided by industry sales— as the standard for
identifying high-tech industries. Under this methodology, industries with high R&D
intensity such as Medical, precision & optical instruments (ISIC/NACE 33),
Pharmaceuticals (ISIC 2423 / NACE 2441 & 2442),Radio, television &
communication equipment (ISIC/NACE 32), Office, accounting & computing
machinery (ISIC/NACE 30),Aircraft & spacecraft (ISIC/NACE 353),and
Management activities of holding companies (ISIC/NACE 7415), belong to the high
tech sectors. These industries create value mainly with its immense ability to innovate
and invent— a classic case of “knowledge economy.”
A known trait for high-tech product demand is its sensibility to
macroeconomic environment. From demand side factors, shrinking purchasing power
from its end-user in a weak economy will impact high-tech industry especially hard
due to its “luxury good” nature. From supply side, reduced ability for manufacturer to
fund necessary production activities in a less liquid capital market will curb the
manufacturer from producing its optimal quantity.
Samuelson and Nordhaus(2001)defined income elasticity as “the
percentage change in the quantity demanded to the percentage change in disposable
income.” It is a measure to capture the degree of impact on consumption of a certain
product when income level changes. For a given demand function Q(I, P) of a good,
the income elasticity, , can be calculated as :
.
The greater the elasticity, the greater the effect income change will have on
the consumption of this product. For goods with income elasticity equal to 1, the
“normal goods”, 1 percent change will bring about exactly 1 percent change in the
quantity demanded. For “inferior goods”, those with income elasticity smaller than 1,
1 percent change in income will only produce less than 1 percent change in
consumption, meaning that as income increase, consumers demand less of that
product. On the other hand, some goods, the “luxury goods”, have income elasticity
greater than 1. This implies that during boom periods, people consume more of this
type of goods as a result of increase in personal wealth. However, as economy
weakens, the drop of demand of this type of goods will be greater than the drop of
income. High-tech products such as consumer electronics or sophisticated machinery
are usually treated as luxury goods with income elasticity greater than 1 in economic
analysis. Therefore, many high-tech firms experience cyclical sales performance,
outperform ordinary manufacturing companies in peak and underperform at valley.
Another reason high-tech industries is especially prone to changes in
business environment relates to the company’s cost of capital and ability to obtain
capital. For example, higher interest rate not only tightens the company’s ability to
invest and pursue future growth, but also affects company’s cost of inventory and in
terms changes their purchasing behavior. In other words, in addition to having a
negative impact on both direct and derived demand, economic downturn will also
harm supplier.
Himmelberg (1994) showed that the rate of technology acquisition has a
statistically significant relationship with capital market performance with data from
179 firms in high-tech sectors. He implied that if external funding becomes scarce,
smaller companies have to rely on internal finance, which may not be as affluent as
other liquidity sources. Without sufficient investment in R&D, the key production
input of their products, will not be enough to support fast business growth and
expansion. The dismal outlook of capital market will disrupt high-tech companies
from normal course of production more than other firms in other industries.
Hax and Candea(1984) proposed an Economic Order Quantity (EOQ)
model that help firms to estimate the optimal purchasing level with minimum annual
total cost. This model should be relevant in discussing sales volume in high-tech
industry because a huge portion of the demand is derived demand—the customer of a
manufacture in the high-techs sector will most likely be selling to another
down-stream manufacturer who uses the components he purchased to assemble or
produce something closer to the consumer end. In other words, the demand in
high-tech industry in fact largely comes from inventory purchase, and will be
influenced by the factors that decide the buying firms purchase quantity—making
EOQ a relevant discussion here. EOQ model assumes that
A. demand is continuous at a constant rate,
B. the inventory process continues indefinitely,
C. quantities ordered, storage capacity and available capital are without
constraints,
D. replenishment is instantaneous,
E. costs are time invariant,
F. no shortages are allowed,
G. quantity discounts are not available.
Under these assumptions, the economic ordering quantity refers to the
quantity that minimizes annual total cost including holding cost, ordering cost, and
carrying cost— can be obtained at
where A represents the fixed ordering cost; D the fixed demand rate; r
the inventory carrying cost; and c the unit procurement cost. Following researchers
such as Goyal (1985), Chand and Ward (1987),Aggarwal and Jaggi (1995) have
probed further into this relationship. Some even made further adjustment to carrying
cost, r. For example, Berling (2008), argued the use of interest rate combined with
expected price decrease. However, the relationship between current interest rate and
purchasing quantity is yet clear-cut. For one thing, higher interest rate could imply
smaller purchasing volume if the purchasing firm has to borrow for to pay and thus
raising carrying cost. This would results in a negative correlation between purchase
quantity and interest rate. However, in business practice nowadays, it is common that
the payment to supplier occurred at a later period after the receipt of materials without
charging interest. As a result, this transaction can actually be viewed as an
interest-free loan from the suppler to the purchaser. Thus, all else being equal, a
positive correlation between interest rate and purchase quantity should be observed.
The higher the interest rate, the larger quantity a purchasing firm should order to
exploit this benefit. The final effect depends on the difference of interest cost and
benefit that varies across firms.
In conclusion, because of the high income elasticity of the end-product,
high capital expenditure need, and the derived demand considerations of purchasing
behaviors in the high-tech macro-economic factors such as GDP growth, inflation rate
and interest environment, along with seasonal effects, are drivers one should consider
in formulating a forecast model.
2.2 Industrial Product Forecast
As mentioned in the first section, the major proportion of demand in
high-tech industry is derived demand composed of material purchase for manufacture
of end product. That is to say, compared with consumer products that have direct
linkage to demand and preference, the forecasting of industrial product will have a
different focus that need to factor the change of industry and firm behavior into
account; it will also take more effort to reach accuracy because of the distance
between the producers and end demand.
Dalrymple (1987) found that in practice, firms producing industrial product
prefer the use of sales force composite, a forecasting method of summing the sales
expect of every member of the sales team on their respective responsible regions to
estimate total future demand, is more heavily used compared with forecasting
conducted by companies that sell consumer products. Usually, industrial products
have a more concentrated client base; in other words, the customer pool is composed
of a limited amount of key accounts rather than the mass general consumers in
consumer product. This concentration has raised industrial product’s reliance on
customers because the lost of a single customer can translate into huge sum of loss in
revenue compared with the marginal loss of an individual consumer in consumer
product business. It has also made relationship management to these key accounts
more important. As a result, sales representatives become the person closest to
demand and understand the situations of customers’ best.
On the other hand, Herbig et al (1994) also found in his survey on 150
companies that compared with their consumer product counterparts, companies in
industrial product industry will allocate more focus on industrial analysis and the
trend of the market. Herbig, who uses high-tech industry to illustrate this point, stated
that in industries where technological innovation plays a heavy role in shaping
consumer behaviors, the future demand of existing products highly volatile and
uncertain. As a result, companies are more concerned about the competitive situations
and relative competitive advantage across different players in the market because this
would be the determinant for their sales volume in the long run. They will also
conduct analysis on industry trend to capture the potential technology transition,
which is less common in consumer products.
Because of the complicated market structure and uncertain product life, the
accuracy for industrial product is generally perceived to be lower than consumer
product by both researchers and management in practice. Dalrymple (1975)
conducted a survey on 175 representative companies to study the forecast practice
among different industries. The forecasting methods of these firms were diverse and
varied substantially : qualitative measures included executive opinion, leading index,
and life cycle analysis; quantitative methods ranged from simulation, diffusion index,
exponential smoothing, moving average, regression, input-output model…etc; mixed
methods like sales force composite, intention-to-buy survey, trend projection were
also adopted. Among these miscellaneous means of forecasting, nevertheless, the
error rate for industrial product is significantly higher than consumer-oriented firms,
the former being 7.65 percent and the latter 6.7 percent. Also, Dalrymple’s research
indicated the wider the operation a company has, the greater error rate this company
will encounter. Combined with the discussion in the beginning of this chapter, this
implies that high-tech manufacturing firms targeting at international industrial
companies will experience a higher than average error rate in forecasting.
Concluding from the above literature, future demand for industrial
products depends largely on a selected few customers, involves a wider range of
locations, and faces high uncertainty due to fast-changing technology. The forecasting
results generally come with higher error rates and rely more on sales personnel
understanding of the region or segment.
2.3 Heterogeneity Among Purchasers
As mentioned in the previous section, customer pool for industrial product
companies are relatively concentrated and individual contribution of a key account
could be significant to the total sales. Naturally, one would think that a thorough study
on these major customers should reveal crucial information for prediction of future
purchasing buying behavior. It would also be quite feasible since most of the buyers
are companies with global presence and data can be obtained through direct public
information collection or indirect estimation from other second-sources. Therefore,
one reasonable and cost-effective method to improve forecasting results would be to
incorporate purchaser behavior analysis into regular forecasting.
For consumer behavior analysis, one usually would take out demographic
variables such as gender, age, or nationality…etc—corporate equivalents of these
demographic attributes would be its different business models. Various aspects of
business model involve an enterprise’s role in the value chain, its main clients and
suppliers, operational regions, and competitive strategies…etc. A study into business
model will help the company to understand the different factors driving the
purchasing behavior of each customer and in turn achieve results similar to customer
behavior analysis that eventually can be utilized to improve customer retention and
assist the company in further develop sales promotion strategy.
In recent years, business models have undergone several waves of transition
due to the intensifying focus on specialization in each link of the high-tech industry,
the economic transformations in developing countries, and awareness for brand value.
In the 1980s, major high-tech companies in developed countries conducted a series of
vertical disintegration and outsourcing in an attempt to focus on their core
competency of R&D and reducing manufacturing cost. This change has brought about
the formation of OEMs (Original Equipment Manufacturer) in developing countries in
south eastern Asia. These companies are purely manufacturing companies that accept
orders from these high-tech companies and do not own the capability to design their
own product. In 1990s, further emphasis on operation streamlining from leading
high-tech companies and heated competition among OEMs prompted ODMs
(Original Design Manufacturer) to come into existence. ODMs tried to differentiate
themselves from OEMs by providing one-stop shopping services that cover design,
procurement and manufacturing of the end product. Their services cater to the specific
needs of each client, and these leading technology firms at the top of the value chain
are only responsible for the marketing of the product, which, in fact seize the most
value proposition of total profit. Now, as margin begins to dwindle, some of these
OEMs or ODMs have strived to move upward the value chain by venturing into
private brand, and started to sell their products to Europe and America. For one thing,
these various forms of business model can affect factors in common practice of sales
forecast for industrial product such as competitive analysis and industrial trend. More
importantly, they also signify different company attributes such as their respective end
consumer segments or ability to have supernormal earnings as an influence of brand
value that would eventually reflect in different purchasing patterns.
While consumer heterogeneity has been a major issue in marketing science
and a proliferation of literature on various products, industries can be found; yet,
heterogeneity in purchasing firms has remain largely untapped and is to be addressed
more in literature. More research effort should be dedicated to make use of the
information possible to be yielded with customer features.
2.4 Multi-level Forecast Model
Chapter 2 reviews the basic dynamics of high-tech industry and past
forecast literature relating to this sector. Specifically, from both demand and supply
aspect, high-tech industry is prone to macro-economic environment for several factors,
such as high income elasticity of the end product, and sensitivity to interest rate due to
funding constraint and inventory cost. Also, the huge portion of high-tech industry
output are industrial products with markets facing derived demand. Forecasting for
these products differ from that of consumer products in that buyers are more
concentrated, product life cycles are shorter and contingent to industrial development,
and with higher error rate. However, since the sales volume of a company largely
depends on a number of key accounts, it should be useful to utilize purchaser specific
attributes to further capture variation in demand.
The conclusion leads to the use of a model combining both
macro-economic factors affecting general quantity demanded and the use of
respective customer features to explain dissimilarities across key accounts of a certain
company. In other words, beside projecting future demand with a top-down approach
using economical variables that impact sales of the industry, this forecast model must
also take the heterogeneity among its major customers to consideration so as to
estimate the respective variation in purchasing quantity each customer might have
given its individual business model and the behavior under the business model it
adopts. By implementing a two-stage model to fine-tune for specific variations that a
general model fails to account for, the aggregate forecast result can expect to be more
accurate.
Although current literature lack similar application on industrial product,
similar research came be found on consumer products. One research by Lo et al.
(2008) on LCD monitor market first used variables related to macro-economic
situations to generate forecast for the LCD market as a whole and then fine-tuned the
result with different product attributes in this category. The model proposed involved
a hierarchical forecasting with 3 levels of model showed below. It first projected the
quantity demanded with economic variables such as GDP per capita in major markets
and interest rate; then each product group were divided by a certain attribute such as
resolution or monitor size and a corresponding model was constructed for each
category; lastly, the forecasts results were applied to different geographical regions
and combined to produced the final estimation. This hierarchical model that pooled
together factors in external environment and specific market attribute can be further
extended in the forecast of industrial product.
Chapter 3 Research Method
The research methods employed in empirical analysis include two
parameter estimation methods and one simulation algorithm. For model parameter
estimation, we compare the traditional Maximum Likelihood Estimation method with
Hierarchical Bayesian method to find out the strengths and weaknesses of the two.
The purpose is to propose an estimating method with higher reliability and better
accuracy. In the case of Hierarchical Bayesian model, Markov Chain Monte-Carlo
simulation is used to generate the parameters of posterior probability distributions.
3.1 Hierarchical Bayesian Method
Classical statistics developed estimate based on sampling theory, which
generates the parameter estimation purely based on the sample data. In contrast,
Bayesian statistic method is developed from Bayes’ Theorem. The parameter
estimation process involves combining sample observations with prior information to
come up with the posterior probability distribution for the parameter. The estimate is
then generated while minimizing the expected loss.
Although Bayesian estimates are not unbiased, but they satisfy consistency
because they are generated while minimizing estimation error and mean square loss.
also satisfy sufficiency. Therefore, Bayesian estimates still have good properties.
Furthermore, one can compute posterior probability density function for parameters
with Bayesian method to bring additional insight into analysis.
While Bayesian statistics overcomes the problem of high estimation error
with classical statistics under limited sample size, Hierarchical Bayesian method
further takes into account the “unobservable heterogeneity” between each subject and
uses it to modify the projected values respectively. As such, Hierarchical Bayesian
method can offer an even more accurate estimation than Bayesian method. Compared
with Bayesian method that makes assumption of one prior probability distribution,
Hierarchical Bayesian method makes multiple levels of assumptions on prior
probability distribution—the first level is aimed at capturing the variation in a single
subject, whereas further levels will probe into the unobservable heterogeneity
contributed to these differences. This process is done by the construction of multiple
stages models. The “within-subject” model is a linear regression model that leads to a
matrix of individual-level regression coefficients. The independent variables in this
level are general variables used for explaining the differences in observations for a
subject. In the later stages, “between-subject” models come into play in the form of
multivariate regression models that portray the relation between the coefficients from
the previous level with subject heterogeneity. By doing so, Hierarchical Bayesian
method will allow for “multiple sources of uncertainty” (Lenk, 2001) when explaining
variation in observed samples. Managers and researchers alike will benefit from the
additional information brought by the use of multiple-stage models.
With the feature mentioned above, Hierarchical Bayesian method not only
allow forecast in both aggregate level and individual subject level, but actually has a
mechanism to modify the estimate itself when additional posterior information. As the
sample of a subject increases, the parameter estimate will be influenced by its
heterogeneity—i.e. its posterior probability distribution—more than it is by the
within-subject estimation—the prior probability distribution that represent the first
level estimation. This self-modifying mechanism will be especially useful in
forecasting with limited values, because it at most gives estimation close to the prior
probability distribution. Therefore, it is even more accurate than simple Bayesian
method.
Researchers have derived and proved that the posterior mean for a
Hierarchical Bayesian parameter estimate is actually a convex combination of the
within-subject MLE (M) estimate and the between-subject estimate. That is to say, the
bayes estimator can adjust for the original MLE estimates between the two depending
on their relative accuracy. If more subjects are involved in the estimation process,
more weight would be put on within-subject estimator and the Bayes estimator will
eventually tilt toward a value near MLE estimator. In other words, Bayes estimator
will not differ much from MLE estimator in large samples; its value is better
distinguished when samples are limited.
3.2 Markov Chain Monte-Carlo Methodology
Markov Chain Monte-Carlo Methodology is an extension of Bayesian
estimation. It treats the joint posterior distribution as the target distribution and
simulates it with Markov Chain. Then, Gibbs sampling will be employed to obtain
estimation for parameters by iteration. Gibbs sampling will use the previous sampling
to condition later samplings, and current sampling result is only affected by the
previous sampling, which is in line with the characteristics of Markov Chain. The
iteration process takes the previous sampling result as the given condition of current
sampling that enters into the conditional distribution. This repeating sampling process
with updated conditional posterior distribution will help the result to approximate the
real posterior distribution.
Chapter 4 Empirical Analysis
4.1 Sample Summary
The following empirical analysis utilizes data drawn from the sales record
of a leading IC design house in Taiwan. The time range of the record ranged from
January, 2006, to March, 2009, a total of 39 periods of monthly sales quantity. In this
product category, the company has acquired 40 customers by the end of the record
period. These customers are generally mobile phone manufacturing companies that
either produces mobile phone under their own brand or companies providing EMS
(electronic manufacturing service) to international mobile phone companies that wish
to streamline their process. These mobile phone makers purchase chipsets used in
handsets from the IC design company and assemble it to their products along with
other components to make a complete cell-phone. In other words, the product of this
IC design company does not share directly link to mobile phone end-users; rather, it is
offered to derived demand from mobile phones makers.
Since the beginning of this data set, the company has collect another set of
information of the top 8 accounts, which accounted for 60% of the sales volume in
year 2006. The information included is listed below. Segmentation is used to
distinguish the business model of the key accounts itself, including pure IC design
House, ODMs, and mobile phone companies with their own brands. On the other
hand, sales mode refers to the customers of different business models of these key
accounts, including mobile phone companies (both in China and overseas), mobile
phone operators (both in China and overseas), PCBA (Printed Circuit Board Assembly,
meaning that the company process the chipset in to PCBA and sell it to medium/small
size mobile phone manufacturers), and Open BOM (Browser Object Model, offering
design service for major mobile phone companies). Sales region refers to the location
of the customers of the key accounts.
The summary statistics of the sales quantities of the 8 accounts are listed
below. Considering these 8 accounts altogether, sales quantity has a positive skewness,
which is to say that the observation shows a tendency to be smaller than mean. Also,
sales volume of these 8 accounts, plotted in Graph 2, showed high variation across
month. When constructing a forecast model, one should be careful in the possible
seasonality issue and include variables that might mitigate this effect.
Exhibit 1 Account Information
Customer Attributes Categories
Segmentation IC design house, ODM, Brand
Is ERP/MRP employed? Yes; No
When is ERP/MRP employed Date
Inventory Policy Days of Inventory
Publication Time of Financial Report Date
Shipment Information
Sales mode (%) China Brand China Operator
PCBA Open BOM
Overseas Sales Region (%)
China Russia Taiwan
India
South Eastern America Latin America
ROW
Source: company data
.
Exhibit 2 Descriptive Statistics of 8 Accounts (sales quantity in thousands)
Account N Minimum Maximum Sum Mean (Std. Deviation)
Std.
Deviation Variance Skewness Kurtosis A87 39 204 2299 37843 970.333 546.254 298393.070 0.740 -0.304
(87.471)
A84 39 81 967 18009 461.769 208.869 43626.340 0.593 0.119
(33.446)
A110 39 64 1615 25324 649.333 368.147 135532.439 0.693 0.007
(58.951)
A2 39 45 1734 23457 601.462 414.087 171468.255 0.915 0.050
(66.307)
A80 39 102 587 11965 306.795 133.537 17832.167 0.492 -0.540
(21.383)
A14 39 39 784 9846 252.462 149.194 22258.992 1.303 2.875
(23.890)
A29 39 50 774 12354 316.769 149.956 22486.866 0.593 0.901
(24.012)
A122 39 18 1255 16425 421.154 300.258 90155.081 0.800 0.283
(48.080)
Total 39 1192 8431 155223 3980.077 1753.839 3075950.757 0.474 -0.215
(280.839)
Valid N 39
Graph 1 Monthly Sales Quantity (in thousands)
4.2 Model Construction
As discussed previously, the HB model will involve 2 stage of analysis—
the first stage will be a structural form model using macro-economic variables as
explanatory variables to represent the general factors influencing sales of each
customer. Then, the second stage will draw insight from a model that incorporates
firm specific features to allow adjustment for each account’s heterogeneity. For the
purpose of out-sample forecasting comparison, 3 periods of observation will be
reserved and 36 periods will be used for parameter estimation. Data for aggregate
level were extracted from TEJ database (Taiwan Economic Journal); for individual
level, firm attributes in Exhibit 1 are used as explanatory variables of the model.
In both the aggregate and individual stage, step-wise regression procedure
is adopted to choose variables with the maximum collective explanatory power from
the pool of relevant factors initially proposed in Exhibit3. By doing so, the number
of variables will be reduced since variables with less contribution will be screened
out. Only those with highest relevancy will be retained. As noted in later sections,
most of the sales of end product are directed toward China region. Consequently,
macro-economic situations in China would be the most relevant issue in the
aggregate level. Therefore, it makes more sense that many macro-economic indices
entertained in the model are China-based statistics instead of the world. Also,
industry specifics such as demand/supply conditions and compliments demand are
considered to relate to the characteristics of technological product forecasts, which
emphasize on industrial competitive and trend analysis. For example, BB ratio is an
indicator of demand/ supply strengths in semi-conductor industry. Dividing total
industry order quantity with number of orders filled, it measures the degree of
demand exceeds current supply capacity. A BB ratio over one signifies greater
demand than current supply, whereas a BB ratio less than one imply oversupply in
the industry.
Exhibit 3 Initial Macro-economic Variable Pool
Category Variable
Capital Market
NumStock :the numer of stock that company has issued
StockA :A Stock index in China
StockB :B Stock index in China
LoanRate :current one-year borrow rate in China
AvgRate:current one-year average rate in China
DIProj :numbers of direct investment projects in China
International Trade Exchage Rate :RMB to TWD
Price Index
CPI:consumer price index in China
CCI:consumer confidence index in China
PMI:purchase manager index in China
Infrustructure
TeleServ :Telecom Service Availability in China
TeleIm :Telecom Equipment & Services Import In China
TeleEx:Telecom Equipment & Services Export In China
Supply/ Demand/
Complements
BBRatio:Book-to-Bill ratio
NumPhone :industry sales volume of mobile phone in China
LCD:Small Size LCD shipment
Economic indices
Leading :Leading Macro-Economic Climate Index in China
Concur:Coincident Macro-Economic Climate Index in China
Warning :China Monitoring Signals of Macro-Economic Climate Index in China
Exhibit 4 Aggregate Variables
Category Code Definition
Seasonality Yt-12 sales quantity lag 12 periods
Macro-economic
LoanRate China Official Interest Rates of Loans (1 year)
Concur China Macro-Economic Climate Index Warning China Monitoring Signals of
Macro-Economic Climate Index
PMI Purchase Manager Index
Demand/Supply BBRatio Book-to-Bill ratio
Complements lnLCD Small Size LCD shipment; taking natural logarithms
The first stage aggregate model was derived in the form of multiple
regression and can be written as
Because a lagged 12 period variable is included, the original data size of 36 has been
reduced to a size of 24 effective samples. Therefore, the dependent variable, , is a
vector, representing the 24 period monthly sales quantity of account .
is a matrix, composed of the values of 7 selected economic variables
(listed in Exhibit 4, including intercept) in the 24 periods. is a vector of
regression coefficients obtained by regressing the sales quantities of account on
these 8 explanatory variables. Finally, is the normally distributed error term
vector. The prior probability density function for the variance is an Inverted
Gamma Distribution written as
In the individual level, will be treated as the dependent variable that
would be influenced by firm specific attributes. In order to compare forecast results,
the MLE method is also employed to generate parameter estimation. The MLE
estimation of would be
Exhibit 5 Correlation Matrix of Aggregate Variable and Sales
Sales LoanRate PMI BBRatio LCD Concur Warning
Sales 1.000 0.705 -0.265 -0.819 0.908 0.098 0.062
0.000a 0.118 0.000 0.000 0.570 0.718 LoanRate 0.705 1.000 0.205 -0.765 0.669 0.680 0.680
0.000 0.231 0.000 0.000 0.000 0.000 PMI -0.265 0.205 1.000 0.133 -0.345 0.705 0.705
0.118 0.231 0.438 0.039 0.000 0.000 BBRatio -0.819 -0.765 0.133 1.000 -0.780 -0.280 -0.314
0.000 0.000 0.438 0.000 0.098 0.063 LCD 0.908 0.669 -0.345 -0.780 1.000 0.044 0.042 0.000 0.000 0.039 0.000 0.800 0.808
Concur 0.098 0.680 0.705 -0.280 0.044 1.000 0.942
0.570 0.000 0.000 0.098 0.800 0.000
Warning 0.062 0.680 0.705 -0.314 0.042 0.942 1.000
0.718 0.000 0.000 0.063 0.808 0.000 a. P-Value
At the aggregate level, loan rate has a positive correlation with sales
quantity, suggesting that firms would rather view storing inventory as an alternative
way of funding similar to an interest rate free loan from the seller. Thus, the higher
the current interest rate is, the more then would order since it become relative cheap
to stock materials. On the other hand, BB ratio shows a negative correlation with
sales. This could be interpreted as an adjustment process of supply and demand. If
current BB ratio is high, meaning order has exceeded available capacity, sales would
decline. This is either done with a price hike from the suppliers or purchasers would
have satisfied the necessary level of stock and reduced ordering. On the other hand,
shipmentof complement goods of mobile phone chipset, small size LCD that
functions as the screen for handsets, has a strong positive correlation with sales. This
is in line with economic intuition that the more chipsets are sold to make mobile
phones, the more LCDs are needed to go with it.
The variables selected here can be categorized into 4 types of indices. A
lagged variable is included to address seasonality problem, which can be observed
clearly from the pattern of monthly sales data. Then, 4 economic factors are deemed
to be effective in influencing aggregate sales volume, namely interest rate, 2
economic indices, and 1 price index. Because these customers are manufacturing
purchasing product as inventory, PMI is in place of CPI (Consumer Price Index).
Book-to-Bill ratio is the measure in semi-conductor industry, which is the amount
booked divided by the amount billed. It reflects the relative strength in demand as
compared with supply. A Book-to-Bill ratio larger than 1 signifies the industry has a
demand stronger than capacity, a sing welcomed by the producers. Reversely, when
Book-to-Bill ratio falls below 1, the producers are essentially over-supplying and
either their price will suffer or the average cost will rise due to slack capacity.
Exhibit 6 Individual Variables
Category Code Definition
Business Model Segment Dummy variable; Brand=1, otherwise=o
Inventory Policy Inventory Days of Inventory
Sales Region
China
percentage of end-product sales allocated the area
Russia India Latin AM
The second stage model regresses the aggregate coefficients on account
attributes to adjust for individual heterogeneity for each customer. It can be
expressed as
is the vector of covariates for account and is a matrix of
regression coefficients because 6 variables representing individual account
heterogeneity are selected in this level (including intercept). here is a
vector of error terms. Note that the smaller covariance matrix of , , is, the
more variation in individual account from the predicted value of aggregate level is
explained by firm attributes. Its prior probability density function is an Inverted
Gamma Distribution
3 categories of account attributes have entered into the second stage
model, namely type of business model, inventory policy, and sales region. In order
to verify on the value-adding effect of branding generally claimed, a dummy
variable “Segment” is created. Accounts operating under its own brand were given a
value of 1 to distinguish it from other business models such as ODM or IDH, in
which cases a value of 0 were assigned. Inventory policy refers to the days of
inventory stocked before this customer sells it to its clients. In addition, different
composition of sales region mix is also a relevant issue in the individual level. The
percentage of final product sold in each region of the world, most notably the
“BRICS”, are shown to influence the quantity demanded by the key accounts. It is
the geometric average percentage across the data period.
4.3 Results & Forecast
This section will present the empirical results with the model mentioned
previously, including preliminary data examination, parameter estimation and also
forecast comparison between the two methods.
Summary statistics for the explanatory variables in both aggregate and
individual level are given below. In accordance with sales pattern, macro economic
variables have a high degree of variation with large standard deviation and range.
For example, among these 36 periods, China Monitoring Signals of
Macro-Economic Climate Index (Warning) can be as low as 78.7 to as high as 121.3,
reflecting the economic cycles experienced during the observed time. Naturally,
business cycles will lower forecast accuracy; however, it might as well be improved
with additional insight into the second level, which would address different
sensitivities of each account to the general environment.
Exhibit 7 Summary Statistics for X
Variable Mean Std.
Deviation Minimum Maximum
Yt-12 5.757 0.830 2.890 7.740
LoanRate 6.911 0.640 5.310 7.470
Concur 101.917 1.882 95.510 103.420
Warning 109.717 12.256 78.700 121.300
PMI 52.596 5.007 38.800 59.200
BB Ratio 0.873 0.082 0.700 0.998
lnLCD 14.790 0.404 13.982 15.507
The second stage would regress on individual firm attributes, of which summary statistics are listed in the following Exhibit 8.
One thing worth pointing out in firm attributes is the importance of sales
to China. Each of the 8 accounts have their own specialized regions; yet at least 20%
of their sales has to do with China, with the most focused one devoted 80% of the
sales in serving customers at that region. Latin America would be another major
region that has strong demand for the end product. One would reasonably expect the
tremendous impact the demand from China could have on their sales volume, and in
turn, the influence on these companies’ purchase from the company.
Exhibit 8 Summary Statistics for Z
Variable Mean Std.
Deviation Minimum Maximum
Inventory 28.750 8.763 15.000 45.000
China 0.563 0.205 20.0% 79.5%
Russia 0.032 0.036 0.0% 10.0%
India 0.106 0.109 0.0% 30.0%
Latin America 0.058 0.085 0.0% 25.0%
The breakdown of account attributes are given below. Once again, it can
be observed that none of the regions possess the same degree of influence compared
with China. Also, notice that these 8 largest accounts all have their sales mainly
directed to the emergent markets. Sales to the “BRIC” region have contributed at
least 50% of their total sales. One may refer that this is a reflection of the high
potential of these developing economies where demands are not fully satisfied and
populations are huge. Nevertheless, one should bear in mind that as mentioned in
literature review, common practice for business intelligence collection in industrial
product business—including the company discuss in this research— is to rely on
sales force composite. In other words, these figures are estimation of sales personnel
of their respective accounts/regions and may suffer distortion from lack of factual
base.
Exhibit 9 Attribute Information
Account Segment Inventory China Russia India LatinAM Sum
A87 0 30 78.0% 0.0% 9.0% 1.5% 88.5%
A84 1 30 60.0% 3.0% 5.0% 0.0% 68.0%
A110 0 30 65.0% 3.0% 12.0% 10.0% 90.0%
A2 1 30 50.0% 0.0% 0.0% 0.0% 50.0%
A80 1 45 20.0% 10.0% 0.0% 25.0% 55.0%
A14 0 15 35.0% 7.0% 30.0% 0.0% 72.0%
A29 1 20 62.5% 1.0% 23.5% 5.0% 92.0%
A122 1 30 79.5% 1.5% 5.0% 5.0% 91.0%
The estimated coefficients are presented in the following pages. Because
the inclusion of lagged 12-period value, the actual observations used for estimation
were only 24 data points extracted from the original pool. As a result, fewer
estimates can reach a statistically significant t-value. For the second stage, only 8
observations were used to estimate , making it even harder to demonstrate
statistical significance. Nonetheless, the covariance matrix of , , showed
extremely small values, showing effectiveness of the estimation.
The coefficient matrix showed results somewhat different from that of
initial correlation matrix analysis, and this has once again justified the need to
construct a second level model to fine-tune for individual heterogeneity. For
example, while LoanRate shared a positive correlation with sales using the
summative data, the coefficient of it for some accounts are negative. For similarly
statistically significant level in lagged value Yt-12, A2 has a positive value while
A80 a negative one. Even variables with the same signs across account, such as BB
ratio, have a difference in the degree of influence. Clearly there is a need to further
probe into individual factors contributing to these differences.
A look into the second level would show some enhancing or mitigating
effects the firm specific attributes have on the aggregate level coefficient. These firm
attributes either increase or decrease the importance of the macro-variables in
explaining sales quantity. For example, the dummy variable, “Segment”, has a
coefficient of 0.616 on BB ratio. Given the fact that BB ratio showed negative
contribution to sales quantity, firms having a Segment value equals to 1— namely,
those operate under their own brand – will react less negatively to high BB ratio.
One may as well say that firms with their own brands are more resistant to the
dynamic adjustment of demand/supply of the industry and will have a smoother
sales pattern compared with those adopting non-branding business models.
On the other hand, one can also see that a negative reaction to BB ratio
will be intensified if the account puts its emphasis on some geographical regions.
With a coefficient of -5.500, 1 more percentage of end-product sales allocated in
the China region will affect the coefficient for BB ratio in the aggregate model to
reduce by 5.5%. The same is true for Latin American region—1 more percentage of
end product sold to Latin America will reflected in 3.54% lower coefficient for BB
ratio in aggregate level. One may consider these two regions to be more aggressive
in terms of adjustment to imbalanced supply/demand. Alternatively, one could
reason that the two regions are more sensitive to shock and having clients making
transaction in these areas will make the sales of the company more volatile.
Comparison between A87 and A122 would be a vivid example of the
interplay of firm attributes in determining coefficients in aggregate level. The 2
companies have the highest percentage sold to China and Latin America, with A122
slightly higher in both regions. As a consequence, they are both more vulnerable to
BB ratio. However, coefficient of BB ratio for A122 is slightly lower despite higher
sales in volatile regions. The reason being is that A122 is a branded company, which
can enjoy some mitigating effect to ameliorate the negative impact from
supply/demand side. Interestingly enough, this finding is in line with contemporary
perception that branding will bring more value-added to the company and make its
customers less price sensitive during a time when the market is saturated with
oversupply.
Exhibit 12 presents the covariance matrix of error term, . As mentioned
in previous discussion on estimating methods, one benefit of Bayesian statistics is
effectiveness, which means smaller estimation error. By observing, one can easily
tell that these values are so small as to be negligible. Even though coefficients in
individual level failed to reach statistic significance, an extremely small estimation
matrix still implies that models at aggregate level and individual level combined
have been able to explain for most of the variations in sales quantity.
Exhibit 10 Estimates of
Account Intercept Yt-12 LoanRate PMI Concur Warning BBRatio lnLCD A87 -5.623 0.031 0.084 0.090* 0.131 -0.054 -3.140* 0.143 (7.953)a (0.230) (0.383) (0.036) (0.097) (0.033) (1.243) (0.347)
A84 -3.962 0.050 0.702* 0.033 0.138 -0.044 -1.571 -0.241 (6.795) (0.200) (0.324) (0.036) (0.088) (0.029) (0.938) (0.343)
A110 -5.516 0.062* -0.020 0.043 0.100 -0.029 -2.900* 0.344 (7.768) (0.146) (0.313) (0.036) (0.094) (0.029) (1.112) (0.337)
A2 -3.656 0.464 -0.027 0.005 0.116 -0.019 -0.927 -0.170 (6.661) (0.195) (0.315) (0.037) (0.091) (0.031) (1.120) (0.377)
A80 -3.973 -0.403* 0.690 0.022 0.087 -0.054 -0.629 0.253 (10.974) (0.140) (0.410) (0.036) (0.130) (0.032) (1.582) (0.404)
A14 -4.571 0.365 -0.425 0.025 0.092 -0.008 -0.614 0.104 (6.245) (0.287) (0.383) (0.036) (0.092) (0.032) (1.586) (0.383)
A29 -4.083 0.258 -0.242 0.028 0.063 -0.009 -1.683 0.339 (7.596) (0.253) (0.348) (0.038) (0.098) (0.033) (1.328) (0.374)
A122 -4.428 0.079 0.337 0.047 0.189* -0.071* -2.920* -0.307 (7.424) (0.152) (0.302) (0.036) (0.094) (0.029) (1.123) (0.343)
a. Std. Deviation of estimation
* Statistic significance at 0.1 confidence level
Exhibit 11 Estimates of
Intercept Yt-12 LoanRate PMI Concur Warning BBRatio lnLCD Intercept -2.473 2.149 -2.570 -0.151 0.100 0.154 1.694 -0.481
(8.842)a (3.218) (3.447) (2.998) (3.005) (2.977) (4.955) (3.472) Segment 1.169 -0.043 0.309 -0.017 0.007 -0.003 0.616 -0.171
(6.967) (0.497) (0.582) (0.450) (0.454) (0.454) (1.413) (0.580) Inventory -0.045 -0.050 0.069 0.005 -0.002 -0.004 -0.018 0.020
(0.419) (0.093) (0.100) (0.088) (0.089) (0.088) (0.134) (0.100) China -2.034 -0.485 0.661 0.086 0.156 -0.103 -5.500 -0.338
(9.402) (1.473) (1.656) (1.362) (1.358) (1.354) (3.586) (1.688) Russia -0.157 -2.219 4.892 -0.009 0.615 -0.365 -1.335 -3.938
(10.051) (7.752) (8.066) (7.130) (7.210) (7.128) (9.629) (8.182) India -2.288 -2.314 1.689 0.225 -0.322 -0.126 -0.062 2.558
(9.752) (5.863) (6.193) (5.538) (5.549) (5.488) (8.050) (6.289) LatinAM -0.914 0.340 -3.128 -0.151 -0.092 0.153 -3.540 1.814
(10.031) (5.334) (5.791) (4.982) (5.002) (5.019) (8.296) (5.770)
a. Std. Deviation of estimation
* Statistic significance at 0.1 confidence level
Exhibit 12 Posterior Estimate of
Intercept Yt-12 LoanRate PMI Concur Warning BBRatio lnLCD Intercept 0.372 0.003 0.002 -0.002 0.003 0.000 0.000 -0.003
(0.441)a (0.204) (0.180) (0.244) (0.199) (0.202) (0.187) (0.246) Yt-12 0.003 0.280 -0.014 -0.001 0.001 0.000 0.003 -0.002
(0.204) (0.240) (0.144) (0.177) (0.156) (0.148) (0.152) (0.189) Loan Rate 0.002 -0.014 0.313 0.002 0.004 -0.005 0.003 -0.014
(0.244) (0.177) (0.160) (0.289) (0.170) (0.173) (0.166) (0.207) PMI -0.002 -0.001 0.002 0.263 0.005 -0.001 0.003 -0.003
(0.199) (0.156) (0.149) (0.170) (0.226) (0.143) (0.150) (0.178) Concur 0.003 0.001 0.004 0.005 0.262 -0.001 -0.003 -0.006
(0.202) (0.148) (0.144) (0.173) (0.143) (0.230) (0.150) (0.184) Warning 0.000 0.000 -0.005 -0.001 -0.001 0.265 -0.001 0.004
(0.187) (0.152) (0.144) (0.166) (0.150) (0.150) (0.232) (0.187) BBRatio 0.000 0.003 0.003 0.003 -0.003 -0.001 0.360 0.008
(0.246) (0.189) (0.174) (0.207) (0.178) (0.184) (0.187) (0.369) lnLCD -0.003 -0.002 -0.014 -0.003 -0.006 0.004 0.008 0.302
(0.203) (0.156) (0.154) (0.180) (0.156) (0.163) (0.158) (0.194)
a. Std. Deviation of estimation
* Statistic significance at 0.1 confidence level
Forecasting results are compared against that of MLE method. With MLE
method, there was only 1 level of multiple- regression composed of Macro-economic
variables. It ignores the difference in firm attributes and attempt to forecast sales
quantity solely with variation in aggregate level. The estimates of using MLE are
listed below and used for both in-sample and out-sample tests. For in-sample test, the
fitted values were compared against the 24 periods of observations used to construct
the models. MAPE (Mean Absolute Percentage Error) is calculated for both methods
to evaluate the level of deviation of fitted values to actual values. The formula for
MAPE is
where is the fitted value for period t and the actual value. The same
procedure is taken for out-sample test, which is conducted using the 3 left out periods
from January 2009 to March 2009.
As shown in Graph 3 and the fitted statistics, one can observe that the
forecasting power of the two methods is quite similar. MLE method has slightly
higher R statistics in in-sample test, whereas HB has lower MAPEs of total sales
quantity in both in-sample and out-sample tests.
Exhibit 13 MLE Estimates of
Variable Mean Std Deviation
CS -6.378 18.498
YT_12 0.073 0.301
LoanRate 0.133 0.686
PMI 0.037 0.028
Concur 0.130 0.164
Warning -0.039 0.028
BBRatio -1.796 1.580
lnLCD 0.114 0.500
Exhibit 14 Fitted Statistics
Account In-sample MAPE Out-sample MAPE
HB MLE HB MLE
A87 29.278% 28.929% 481.649% 293.322%
A84 34.616% 27.370% 37.191% 195.834%
A110 26.330% 26.702% 154.665% 261.340%
A2 27.150% 27.552% 42.814% 41.040%
A80 18.295% 17.100% 21.098% 25.727%
A14 20.330% 19.554% 15.343% 9.649%
A29 35.118% 33.543% 65.287% 26.000%
A122 35.750% 35.416% 237.896% 484.872%
Total 15.876% 15.937% 118.300% 135.193%
Multiple R 0.864 0.878
R-Squared 0.747 0.771
Error Std. DEV 0.337 0.321
Graph 2 Forecast Result (8 accounts)
Chapter 5 Conclusion
5.1 Managerial Implications & Suggestions
This study proposed a two stage model to account for both economic influence and
firm specific attributes that contributes to the differences in purchase quantity among different
key accounts. The purpose of this method is to extend the idea of customer heterogeneity
commonly applied in sales forecast for consumer goods to the forecast of industrial products,
where a number of key accounts dominate the majority of the sales volume. In estimating the
parameters of each level, Hierarchical Bayesian method is applied and MLE method is used as
a comparison alternative. When testing the forecast power of the two methods, results are
similar with HB slightly better in out-sample forecasting.
First of all, the method presented in this research poses an opportunity for
management to develop a more customized forecasting procedure that helps them better
capture the heterogeneity among customer, which in turn can improve their estimation of the
aggregated total demand in this ever-changing business environment. For management, low
accuracy in the forecast of high-tech industry has been a problem that is hindering the
effectiveness of decisions such as a more efficient production planning or appropriate pricing,
and both research and survey have shown that companies manufacturing industrial products
tend to conduct sales forecast with different approaches compared with those in consumer
product industry. While consumer product industry strives to fully utilize information regarding
customers’ behaviors to optimize marketing results, same measure is rarely adopted in
industrial products, where macro-economic environment and future industry trend have been
more of the attention. However, this top-down approach requires more subjective judgment
from the management or analysts, whose knowledge and information might be confined.
Therefore, if it is combined with another bottom-up approach that address sales forecast from
the other side, namely the customers, results produced should be more comprehensive and
accurate. Therefore, management might consider modifying their current practice and
developing a measure to record the characteristics of their customers. By borrowing the best
practice from other industries, they might come up with a model that incorporate different
levels of factors affecting their sales and have a better knowledge of future demand.
5.2 Limitation & Future Directions
The most significant limitation of this research is the representativeness of the
coefficients. Although they show economic sense when explained, few of them actually show
statistic significance due to the constraint of limited observations. Therefore, one should be
careful when interpreting the size, relative scale of and . Instead using the model in this
research for future forecast, it would be safer to say that these models are examples of how
client features could be included and help the forecasting quality of the company.
On the other hand, the data of this research is collected from an existing company
which has yet to adjust their forecasting system to account for the different customer attribute.
Nor do they have very solid idea what are the crucial variables with highest explanatory power.
Therefore, when many of the variables contained in the original data do not show high
correlation with sales, there might be a great deal of other variables with the same relevancy
being left out. This is also a common mis-practice in firms adopting customer database where
database failed to collect the most relevant information and responsible personnel could not
produce optimal results without these crucial information and sometimes might even have to
compromise with currently available variables with small explanatory power. In the future, if
information about customer attributes can be further expanded to cover other aspects that might
also be relevant in explaining sales variation in individual account, forecast results should be
improved. For companies wish to implement this forecasting mechanism, this model
misspecification problem resulted from the gap between the forecast practitioners and
information collector should be eliminated. It would be more effective to have thorough
communication between the sales force composite responsible for collecting client information
and the intelligence department who actually do the final modeling. Understanding from each
party could assure the company with higher chance of collecting the “right” information that
would eventually contribute to better forecasting results.
