17

SIMULATION OF THE HIRFINDAHL-HIRSCHMAN INDEX: THE CASE OF

THE ST. LOUIS BANKING GEOGRAPHIC MARKET

ABSTRACT

This paper proposes three types of functional specifications of firm size distribution in an attempt
to simulate the Hirfindahl-Hirschman Index or HHI. The three types of specifications are
arithmetic progression, linear function, and a group of non-linear specifications in which the firm
size is a function of the firm rank. Empirical simulations are conducted for the St. Louis banking
market using non-linear specifications and bank data from 1994 and 2002. Simulation results
show that relatively higher accuracy is achieved for inverse, power, log and quadratic
specifications.

INTRODUCTION

The Hirfindahl-Hirschman Index (HHI) is a standard index used in analyzing the degree of
market concentration of a particular industry in a particular geographic market. The HHI belongs
to a family of indices that also includes the Rosenbluth Index and the Entropy Index (Jacquemin,
1987). All indices measuring concentration utilize percentage shares of individual firms in a
geographic market. The difference resides in how to weigh such percentages (Shepard, 1979).
From this point of view, all indices of distribution and concentration belong to a broad family.
For example, the standard concentration ratio, which calculates cumulative shares of the largest
m firms in an industry with a total n firms, can be seen as a weighted sum of firm market shares,
with weight being 1 for the largest m firms and 0 for the rest of n-m firms. The standard Lorenz
curve involves a comparison between the cumulative market share and cumulative shares of the
number of firms (Devine, et al, 1974). Different indices vary in terms of their emphasis of
different aspects of market structure. For example, while the HHI gives weight to the influence
of large firms, the entropy index tends to emphasize small firms in the shaping of the overall
index. In addition, the Rosenbluth Index includes not only the firm market share, but also the
firm rank. These three indices are similar in that, unlike the Lorenz curve in which the number of
firms does not affect the evaluation of evenness of distribution, the number of firms in an
industry matters. Studies also found strong correlations among different measures of
concentration indices (Nelson, 1963). This has made the selection of different indices dependent
on the ease in collecting data and of calculation. Advances in computation technologies further
reduces the difficulties in handing large data sets, and leaves the matter to the researchers’
personal preference.

The HHI was first used by Hirschman in a trade study in the 1940s where the square root of the
sum of squared market shares were calculated (Hirschman, 1980). In 1950, Herfindahl used the
version of the HHI as it is used today in his Ph.D. dissertation, and in 1959 in a study of the
international copper industry (Herfindahl, 1959). The index came to be known as the Herfindahl
Index after studies by Rosenbluth (1955, 1957). In 1964, Hircshman published a short article in
the American Economic Review claiming original ownership of the index (Hirschman, 1964).
The index has been known by its current name since that time. Since 1982, the Antitrust Division
 18

of the U.S. Department of Justice has been using the HHI as a measure of market concentration
in antitrust issues, along with the four firm concentration ratio. The HHI has also become an
ingredient in constructing some more complex indices for particular types of markets such as an
oligopolistic market (Cabral, 2000).
Although widely used as a measure of industrial structure, so far the HHI for a given year is only
calculated by using that same year’s data. It has not been used to predict and simulate the market
concentration likely to occur in the future. This is largely due to the fact that the market shares
used in calculating an HHI are unpredictable. Market shares are determined by the size
distribution of an industry in a geographic market. For industries that undergo significant
restructuring, it may be particularly difficult to predict the size distribution of the industry.
However, when ample data are available, it may allow a simulation of the size distribution using
data for geographic markets that have numerous firms. The unpredictability due to a few
reorganizations may be overwhelmed or evened out by a large number of units that maintain
existing patterns of behavior. Alternatively, well-specified functional forms may capture the
essential structural relationship that goes beyond a single year. This paper takes a step in this
direction. It specifies the mathematical forms of firm size distribution in a geographic market,
and also conducts empirical simulations using selective specifications proposed. Three types of
specifications are discussed. The first involves specifying firm size distribution as an arithmetic
progression. The second specification assumes that there exists a linear relationship between
firms of different sizes. Finally, a group of specifications are discussed. The commonality within
the group is that all specifications are functions of firm size ranks. The empirical simulations are
conducted by using banking data for the St. Louis metropolitan area.

THE HHI MODEL SPECIFICATIONS

HHIs with even and semi-even distributions

The following gives the definition of the HHI,

HHI1...n = ∑ S n2 (1)
n

where Sn is the market share of firm n where n=1, 2, … and n. The maximum value of the index
is 10,000 when all industry is concentrated in one firm (10,000=100 ×100). The theoretical
values of minimum HHI depend on the distribution of the market shares. Assume a completely
even distribution where all firms in a market are the same size and have the same share of the
1 1
market. With a market having two firms, HHI1..2 = 1002 ( 2 + 2 ) . For a market with three
2 2
1 1 1
firms, HHI1..3 = 1002 ( 2 + 2 + 2 ) . In general with a market having n firms,
3 3 3

1 1002
HHI1...n = 100 × n × 2 =
2
(2)
n n
 19

Equation (2) gives the general formula for a market with an even distribution. Note that if n=1,
then HHI=10,000, the maximum value possible. Now suppose that out of n firms, there are m
dominant firms. The m firm concentration ratio is therefore Stm=S1+S2+..+Sm. For the remaining
n-m firms, market share is 1- Stm Assume first that within both the m dominant firms and the
rest of the n-m firms, there is even distribution. Applying Equation (2) in both groups results in

Stm2 (100 − Stm ) 2

HHI1..n = HHI1...m + HHI m +1...n − m = + (3)
m n−m

In order to avoid a situation where the size of firms in the remaining group is larger than that in
the dominant group, the condition, Stm>m/n, is set. Notice that in Equation (3), if Stm=0, Equation
(3) changes into Equation (2). While Equation (2) describes even distributions of different
numbers of firms, Equation (3) describes semi-even distributions in that between the dominant
and the remaining groups, there are different firm sizes. However, within each group, there is a
uniform firm size. (One exception occurs where Stm=m/n. When this happens, the firm sizes of
both groups are equal.) The reason for dividing the n firms into two separate groups is that
within an industry or a market, there may be a dominant group of firms and the remaining
marginal group of small firms. The two groups may demonstrate different size distributions and
thus require different specifications regarding their size distribution.

Firm market shares as an arithmetic progression

Even and semi-even distributions can be used for approximations either for the dominant or the
remaining group, depending on industries and/or markets. In general, we would expect firms to
have different sizes. Three different size distributions are specified here. The first specification
assumes an arithmetic progression for firms within an industry; the second specification adopts a
linear function that describes size relationships among firms; and the last is a group of
specifications where size is a function of the firm rank.

For an industry where firm market shares constitute an arithmetic progression, firm sizes are S1,
S2, S3, …Sn. Here S2=S1+∆, S3=S2+∆, …Sn=Sn-1+∆, where ∆ is a common difference.
Substituting S2=S1+∆ in S3=S2+∆ leads to S3=S1+2∆. Similar substitution in other terms leads to
S4=S1+3∆, … Sn=S1+(n-1)∆. Since HHI is the sum of the square of these market shares, then

HHI1...n = S12 + S 22 + S32 + ... + S n2

(4)
= S12 + ( S1 + ∆ ) 2 + ( S1 + 2∆ ) 2 + ( S1 + 3∆ ) 2 + ... + ( S1 + (n − 1)∆ ) 2

Expanding the squared terms and collecting relevant terms in Equation (4), gives

HHI1...n = nS12 + ∆S1 (2 + 4 + 6 + ... + 2(n − 1)) + ∆2 (1 + 4 + 9 + ... + (n − 1) 2 )

In the above equation, the second term on the right-hand side contains a sum of an arithmetic
sequence and the last term contains the sum of squares of n-1’s. Appling the relevant summation
formula results in
 20

n(n − 1)(2n − 1) 2
HHI1...n = nS12 + ∆S1n(n − 1) + ∆ (5)
6

Equation (5) gives the HHI for an industry with n firms forming an arithmetic progression in
their market shares. It is a form of uneven distribution not only in that firms have different sizes,
but also that the Lorenz curve for such an industry will show an uneven distribution. The market
share for the first firm is S1/(nS1+(n-1)∆). An even distribution in a Lorenz sense means the
market share is equal to that of the number of firms. That is S1/(nS1+(n-1)∆)=1/n or
nS1=(nS1+(n-1)∆). However, it is easy to see that as long as ∆≠0, there should be nS1≠(nS1+(n-
1)∆). Therefore the assumption of an even distribution is false.

Whether or not there is any industry with such a firm size distribution is an empirical matter. The
advantage of this distribution is that it is easy to calculate and can be used in approximation of a
certain group in the entire distribution of a market.

Firm market shares forming a linear function

This specifies that the sizes of firms in an industry or market form a sequence in which the size
of firm m can be written as a linear function of the size of firm m-1. Specifically, for S1, S2, S3,
…Sn, S2=C+αS1, S3=C+αS2, … and Sn=C+αSn-1, where C is a constant, and α a slope
coefficient. Again, substitute S2=C+αS1 into S3=C+αS2, there is S3=(1+α)C+α2S1. Similar
substitution in other terms results in S4=(1+α+α2)C+α3S1… Sn=(1+α+α2+…αn-2)C+αn-1S1. As
the sum of these market shares, the HHI is

HHI = S12 + S 22 + ...S n2

= S12 + (C + αS1 ) 2 + [(1 + α )C + α 2 S1 ]2 + [(1 + α + α 2 )C + α 3 S1 ]2 + ...
+ [(1 + α + α 2 + α 3 + ... + α n − 2 )C + α n −1S1 ]2

Expanding and collecting the terms results in

HHI1...n = CΣ I + 2CS1Σ II + S12Σ III (6)

where

Σ I = 1 + (1 + α ) 2 + (1 + α + α 2 ) 2 + ... + (1 + α + α 2 + ... + α n − 2 ) 2

Σ II = 1 + (1 + α )α 2 + (1 + α + α 2 )α 3 + ... + (1 + α + α 2 + ... + α n − 2 )α n −1

Σ III = 1 + α 2 + α 4 + α 6 + ... + α 2 ( n −1)

ΣI, ΣII, and ΣIII are functions of α and n and can be reduced to various extents as sums of
geometric progressions. Specifically,
 21

1−α2 2 1−α3 2 1 − α n −1 2 1 − α n −1 2
ΣI = 1 + ( ) +( ) + ... + ( ) = ∑( )
1−α 1−α 1−α n 1−α

1−α2 2 1−α3 3 1 − α n −1 n −1 1 − α n −1 n −1
Σ II = 1 + ( )α + ( )α + ... + ( )α = ∑ ( )α
1−α 1−α 1−α n 1−α

1 − α 2 n −1
Σ III =
1−α2

Equation (6) therefore becomes

1 − α n −1 2 1 − α n −1 n −1 1 − α 2 n −1 2
HHI1...n = C ∑ ( ) + 2C ∑ ( )α S1 + S1 (7)
n 1−α n 1−α 1−α2

The advantage of the arithmetic progression and linear specifications is that both can be written
as a function of the size of one firm. This helps analysis, which looks into how an individual firm
may influence the HHI of a market. A related advantage is to relate HHI to a group of firms. For
example, by definition, the m firm concentration ratio is Stm=S1+S2+..+Sm. From the condition
Sn=(1+α+α2+…αn-2)C+αn-1S1

Stm=S1+(C+αS1)+[(1+α)C+α2S1]+…+[(1+α+α2+…+αm-2)C+αm-1 S1]

That is, Stm=S1(1+α+ α2+α3+…αm-1) + C[1+(1+α)+(1+α+α2)+…(1+α+α2+…+αm-2), or

Stm − C[1 + (1 + α ) + (1 + α + α 2 ) + ... + (1 + α + α 2 + ... + α m − 2 )]

S1 =
1 + α + α 2 + ... + α m −1

If the above equation is substituted into (7) and the relevant terms written as the sums of
geometric progressions, (7) becomes
 22

1 − α n −1 2
HHI1...n = C ∑ ( )
n 1−α
1 − α m −1
S tm − C ∑
1 − α n −1 n −1 1−α
+ 2C ∑ ( )α m
(8)
1−α 1−α m
n
∑m 1 − α
1 − α m −1
S tm − C ∑
1 − α 2 n −1 1−α 2
+ [ m
]
1−α 2 1−α m
∑m 1 − α

Equation (8) gives a hyperbolic relationship between the HHI and m firm concentration ratio,
indicating a positive correlation with complex non-linear characteristics.

Firm market shares as functions of firm size ranks

The distribution of firm sizes can be specified as a function of the size ranks of firms in an
industry or market. The parameters in the functional specifications can be statistically estimated
using existing data. In essence, this method states that there is an inherent relationship between
the size of market share and the rank of the firms, and the later pattern follows the early pattern,
reflected in the constants of parameters estimated.

One such empirical pattern is a variant of the rank-size rule, which is used in urban geography.
The rank-size rule is a simple empirical observation regarding the size distribution of places in
an urban system. It states that given the size of the largest city of an urban system, the size of a
particular place in the hierarchy can be predicted by dividing the rank of the place into the size of
the largest city. That is, the sizes of places are given by a sequence Size1, Size1/2, Size1/3, …
Size1/n, where Size1 is the size of the largest place. This is the original rank-size rule. In this
sequence, dividing each term by the total size of the entire system, there is S1, S1/2, S1/3…S1/n
where S1 is the share of the largest place. That is, the market share of a unit is the relative size of
the largest unit divided by the rank of the unit. Therefore, the HHI of the entire system is

1 1 1
HHI1...n = S12 (1 + 2
+ 2 + .... + 2 ) (9)
2 3 n

In general, there is the inverse model Sn=a+bn-1 where a and b are constants, and can be
statistically estimated. The related HHI is

1 1
HHI1...n = na 2 + 2ab∑ + b 2 ∑ 2 (10).
n n n n
 23

As seen from the above, if a=0 and b=S1, then Sn=a+bn-1=a+S1n-1=S1n-1. The last relationship is
exactly what the rank-size specifies. As a result, (10) becomes (9). In other words, the rank-size
rule is a special case of the inverse function.

In a power form, firm market shares can be specified as a power function of firm ranks. That is
Sn=anb where b<0. The resultant HHI is

HHI1...n = a 2 (1 + 22b + 32b + ... + n 2b ) = a 2 ∑ n 2b (11)

n

When a=S1 and b=−1, (11) turns into (9). This shows a connection between these two non-linear
specifications. The estimation of values of a and b can be obtained statistically by regression
equation LNS=LNa+bLNn, where LN indicates the natural log.

In an exponential function, market share is expressed in a function Sn=aebn and a and b are
constant with b<0. The HHI is

HHI1...n = a 2e 2 (eb + e 2b + e3b + ... + e nb ) = a 2e 2 ∑ e nb (12)

n

The constants a and b can be statistically estimated using LNS=LNa+bn.

Using the logarithmic function Sn=a+bLNn, the squared market share is Sn2=(a+bLNn)2 and the
HHI can be expressed as

HHI1...n = ∑ (a + nLNn) 2
n
(13)
= n(a + 2abLNn!+b 2 ∑ ( LNn) 2 )
2

Finally, the size distribution can be specified as quadratic function Sn=a+bn+cn2 where a, b, and
c are constants to be statistically estimated. The HHI is

n
HHI1...n = na 2 + 2(a + b + ab)(1 + n) + (b 2 + 2c)∑ n 2 + c 2 ∑ n 4 (14).
2 n n

HHI SIMULATIONS USING ST. LOUIS BANKING DATA

This section discusses simulations that adopt a mix of alternative specifications discussed in the
previous section. Equation (3) indicates an HHI as a combination of component indices
calculated for different segments of a market. As stated earlier, the reason for dividing a market
into segments is to capture the behavioral patterns demonstrated by firms of different sizes. For
example, dominant firms in a market may be described by an arithmetic sequence represented by
Equation (5) while the remaining firms may be better described by an even distribution as in
Equation (2), or a linear function series reflected in Equation (7). Alternatively, various non-
 24

linear specifications and their combinations can be considered. The choice of actual
specifications should be based on a study of existing data related to a particular industry.

This paper simulates the HHI for the banking industry in the St. Louis metropolitan area. The St.
Louis metropolitan area contains eight Missouri counties (Crawford (part), Franklin, Jefferson,
Lincoln, St. Charles, St. Louis, and Warren, two Missouri cities (Sullivan city and St. Louis
city), and five Illinois counties (Clinton, Jersey, Madison, Monroe and St. Clair). This area
differs from the St. Louis bank market defined by the Federal Reserve Bank of St. Louis. The St.
Louis bank market includes the city of St. Louis, Missouri counties such as Jefferson, Lincoln,
St. Louis, St. Charles, Warren, and Franklin (excluding the area around the town of Berger), and
Illinois counties such as Madison, St. Clair and Monroe, plus Sugar Creek and Looking Glass
townships in Clinton County. However, as Zhou (1997) discusses, bank deposits in the St. Louis
bank market account for over 90% of the deposits in the St. Louis metropolitan area. Since the
purpose of the simulation is mainly to evaluate the effectiveness of the simulation techniques
proposed, instead of market structure and characteristics of a geographic bank market per se, the
study uses the St. Louis metropolitan area as an approximation of the St. Louis bank market.

Bank deposit data for the St. Louis metropolitan area from 1994 to 2001 demonstrates several
characteristics. First, the largest banking firm accounts for nearly 20% or even higher of bank
deposits, and 40% or above of the HHI in the entire bank market; Second, just under to just over
70% of bank deposits, and 98 to 99% of HHI are concentrated in the top 10% of firms; scatter
plots of market share distribution from 1994 to 2002 consistently depict non-linear curves. These
characteristics show that the crucial element in simulations is to accurately predict the
distribution of the top 10% of the bank market in the St. Louis area, as well as to use non-linear
specification.

Given the characteristics of the existing data, this study adopts the following strategy in
simulation. First, it is assumed that the size of the largest banking firm is known for the future.
This is a reasonable assumption in that the largest banking firm in a large bank market is always
closely watched. Information is readily available to make reasonably accurate estimations
regarding its influence and size in a particular geographic market. Secondly, the simulation will
separate the top 10% of dominant firms from the remaining firms. Essentially, this is to separate
the St. Louis banking industry into three groups in the simulation. The largest banking firm is in
the first group. The rest of the top 10% of the banking firms are in the second group. The
remaining of the banking firms are in the last group. That is

HHI1..n = HHI1 + HHI 2...m + HHI m +1...n (15)

where subscripts indicate the ranks of banking firms. The subscript 2…m implies the rest of the
firms in the top 10% of the banking firms, and m+1…n implies the remaining 90% of the firms.
This means that simulations in this study focus on the firms 2…n.

Thirdly, in choosing specifications, the study adopts various combinations of non-linear

specifications for top 2…m firms and the bottom m+1…n firms. The following table lists these
combinations.
 25

Table 1 Combinations of Model Specifications Used in Simulations

Top 2…m firm model specification

Bottom Inverse Power Log Exponential Quadratic
m+1…n Inverse
firm model Power
specification Log
Exponential
Quadratic

The simulations of the HHI for firms 2…n are contained in the following tables. The headers
indicate combinations of specifications used for the top 10% of firms and the bottom 90% of
firms. For example, inverse-inverse means that the size distribution of firms 2…m of the top
10% of firms is specified as in an inverse function, and so is the bottom 90% of firms.
Alternatively, inverse-exp means that the size distribution of firms 2…m of the top 10% firms is
specified as in an inverse function while that of the bottom firms in an exponential function. The
number shown in the tables are the sum of the HHIs for both the top and bottom groups.
Simulations are conducted for every year from 1994 to 2001.

Table 2. HHI Simulations for n-1 firms

a. Top firms and bottom firms using the same specifications

inverse-inverse exp-exp log-log power-power quadratic-quadratic

1994 460.13 414.33 460.02 556.77 462.86
1995 440.41 393.76 440.85 521.39 444.07
1996 491.89 422.39 484.92 573.95 488.22
1997 594.06 429.92 568.17 660.02 578.56
1998 410.89 288.56 393.69 429.67 404.34
1999 382.32 267.35 368.41 399.95 378.78
2000 383.67 274.24 367.23 395.09 374.84
2001 387.93 307.1 378.96 431.76 384.57

b. Top firms using an inverse function

inverse-exp inverse-log inverse-power inverse-quadratic

1994 458.07 459.45 464.16 459.46
1995 437.63 439.59 443.94 439.81
1996 489.76 491.12 497.35 491.31
1997 592.59 593.66 599.5 593.87
1998 410.25 410.33 421.37 410.46
1999 381.33 381.78 392.83 381.78
2000 382.93 382.93 398.67 382.87
2001 386.81 387.01 405.85 386.99
 26

c. Top firms using an exponential function

exp-inverse exp-log exp-power exp-quadratic

1994 416.39 415.71 420.42 415.72
1995 396.54 395.72 400.07 395.94
1996 424.52 423.75 429.98 423.94
1997 431.39 430.99 436.83 431.2
1998 289.2 288.64 299.68 288.77
1999 268.34 267.8 278.85 267.8
2000 274.98 274.24 289.98 274.18
2001 308.22 307.3 326.14 307.28

d. Top firms using a logarithmic function

log-inverse log-exp log-power log-quadratic

1994 460.7 458.64 464.73 460.03
1995 441.67 438.89 445.2 441.07
1996 485.69 483.56 491.15 485.11
1997 568.57 567.1 574.01 568.38
1998 394.25 393.61 404.73 393.82
1999 368.95 367.96 379.46 368.41
2000 367.97 367.23 382.97 367.17
2001 379.88 378.76 397.8 378.94

e. Top firms using a power function

power-inverse power-exp power-log power-quadratic

1994 552.74 550.68 552.06 552.07
1995 517.86 515.08 517.04 517.26
1996 568.49 566.36 567.72 567.91
1997 654.58 653.11 654.18 654.39
1998 419.19 418.55 418.63 418.76
1999 389.44 388.45 388.9 388.9
2000 380.09 379.35 379.35 379.29
2001 413.84 412.72 412.92 412.9
 27

f. Top firms using a quadratic function

quadratic-inverse quadratic-exp quadratic-log quadratic-power

1994 463.53 461.47 462.85 467.56
1995 444.67 441.89 443.85 448.2
1996 488.8 486.67 488.03 494.26
1997 578.75 577.28 578.35 584.19
1998 404.77 404.13 404.21 415.25
1999 379.32 378.33 378.78 389.83
2000 375.64 374.9 374.9 390.64
2001 385.51 384.39 384.59 403.43

To obtain a sense of the accuracy in the simulations, HHI simulations given in the above tables
and the actual HHI should be compared. Since the purpose of the study is to predict future HHI
using existing data, it is assumed that a simulation based on a previous year’s data is used to
predict the next year’s HHI. That is, a simulation based on the 1994 data is used to predict
1995’s HHI, simulation based on the 1995 data is used to predict 1996’s HHI, … and simulation
based on the 2001 data is used to predict 2002’s HHI. To find the difference between the
simulations and actual HHIs, we subtract actual 1995 HHI from the simulation based on 1994
data, subtract actual 1996 HHI from the simulation based on 1995 data… and subtract actual
2002 HHI from the simulation based on 2001 data. These differences are displayed in the
following tables.

Table 3. Differences between simulations and the actual HHIs

a. Top firms and bottom firms using the same specifications

year of data
used in year to be
simulation simulated inverse-inverse exp-exp log-log power-power quadratic-quadratic
1994 1995 11.69 -34.11 11.58 108.33 14.42
1995 1996 -57.22 -103.87 -56.78 23.76 -53.56
1996 1997 -113.3 -182.8 -120.27 -31.24 -116.97
1997 1998 180.79 16.65 154.9 246.75 165.29
1998 1999 25.1 -97.23 7.9 43.88 18.55
1999 2000 -6.43 -121.4 -20.34 11.2 -9.97
2000 2001 -8.47 -117.9 -24.91 2.95 -17.3
2001 2002 -14.26 -95.09 -23.23 29.57 -17.62
 28

b. Top firms using an inverse function

year of data
used in year to be
simulation simulated inverse-exp inverse-log inverse-power inverse-quadratic
1994 1995 9.63 11.01 15.72 11.02
1995 1996 -60 -58.04 -53.69 -57.82
1996 1997 -115.43 -114.07 -107.84 -113.88
1997 1998 179.32 180.39 186.23 180.6
1998 1999 24.46 24.54 35.58 24.67
1999 2000 -7.42 -6.97 4.08 -6.97
2000 2001 -9.21 -9.21 6.53 -9.27
2001 2002 -15.38 -15.18 3.66 -15.2

c. Top firms using an exponential function

year of data
used in year to be
simulation simulated exp-inverse exp-log exp-power exp-quadratic
1994 1995 -32.05 -32.73 -28.02 -32.72
1995 1996 -101.09 -101.91 -97.56 -101.69
1996 1997 -180.67 -181.44 -175.21 -181.25
1997 1998 18.12 17.72 23.56 17.93
1998 1999 -96.59 -97.15 -86.11 -97.02
1999 2000 -120.41 -120.95 -109.9 -120.95
2000 2001 -117.16 -117.9 -102.16 -117.96
2001 2002 -93.97 -94.89 -76.05 -94.91

d. Top firms using a logarithmic function

year of data
used in year to be
simulation simulated log-inverse log-exp log-power log-quadratic
1994 1995 12.26 10.2 16.29 11.59
1995 1996 -55.96 -58.74 -52.43 -56.56
1996 1997 -119.5 -121.63 -114.04 -120.08
1997 1998 155.3 153.83 160.74 155.11
1998 1999 8.46 7.82 18.94 8.03
1999 2000 -19.8 -20.79 -9.29 -20.34
2000 2001 -24.17 -24.91 -9.17 -24.97
2001 2002 -22.31 -23.43 -4.39 -23.25
 29

e. Top firms using a power function

year of data
used in year to be
simulation simulated power-inverse power-exp power-log power-quadratic
1994 1995 104.3 102.24 103.62 103.63
1995 1996 20.23 17.45 19.41 19.63
1996 1997 -36.7 -38.83 -37.47 -37.28
1997 1998 241.31 239.84 240.91 241.12
1998 1999 33.4 32.76 32.84 32.97
1999 2000 0.69 -0.3 0.15 0.15
2000 2001 -12.05 -12.79 -12.79 -12.85
2001 2002 11.65 10.53 10.73 10.71

f. Top firms using a quadratic function

year of data
used in year to be
simulation simulated quadratic-inverse quadratic-exp quadratic-log quadratic-power
1994 1995 15.09 13.03 14.41 19.12
1995 1996 -52.96 -55.74 -53.78 -49.43
1996 1997 -116.39 -118.52 -117.16 -110.93
1997 1998 165.48 164.01 165.08 170.92
1998 1999 18.98 18.34 18.42 29.46
1999 2000 -9.43 -10.42 -9.97 1.08
2000 2001 -16.5 -17.24 -17.24 -1.5
2001 2002 -16.68 -17.8 -17.6 1.24

Finding the absolute values for numbers in Table 3, and summing up the absolute values along
the columns, results in Table 4, which gives aggregate values of inaccuracy in the simulations.

Table 4 Aggregate Sums of Differences Between Simulations and the Actual HHIs

Top 2…m firm model specification

Bottom Inverse Power Log Exponential Quadratic
m+1…n Inverse 417.26 460.33 417.76 760.06 411.51
firm model Power 413.33 497.68 385.29 698.57 383.68
specification Log 419.41 457.92 419.91 764.69 413.66
Exponential 420.85 454.74 421.35 769.05 515.10
Quadratic 419.43 458.34 419.93 764.43 413.68

The simulation results summarized in Tables 3 and 4 show that simulation errors obtained from
estimations that used an exponential specification are significantly larger than those of other
specifications. These other forms, when used to specify the top firm distribution, along with
 30

various combinations with any other specification, tend to give estimations with smaller
divergence. Table 5 lists simulation errors as percentages of the actual HHIs for different
specifications, averaged over simulations from 1995 to 2002. The average error over a total of
200 simulations is 13.6%. Again, specifications with an exponential form as the top firm size
distribution generally have significantly larger error percentages than those with other forms as
the top firm size distribution.

Table 5. Average Simulation Errors as Percentages of the Actual HHIs (%)

Top 2…m firm model specification

Bottom Inverse Power Log Exponential Quadratic
m+1…n Inverse 11.30 13.32 11.29 21.37 11.12
firm model Power 11.25 14.52 10.33 19.51 10.31
specification Log 11.36 13.25 11.34 21.50 11.18
Exponential 11.39 13.15 11.37 21.61 11.20
Quadratic 11.37 13.26 11.35 21.49 11.18

Of the total 200 simulations, 80, or 40%, have a simulation error within 5% of the actual HHIs;
over half, or 57.5%, have a simulation error within 10% of the actual HHIs. When excluding
simulations with exponential forms as the top firm distribution and considering only those 160
simulations with inverse, power, log and quadratic forms as the top firm distribution, 47.5% of
simulations have a simulation error within 5% of the actual HHIs, and 65.6% have a simulation
error within 10% of the actual HHIs. Table 6 displays simulation errors for alternative
specifications. The higher accuracy from using inverse, power, log and quadratic forms is clear,
especially compared with the poor performance of the exponential form. Of inverse, power, log
and quadratic forms themselves, there is no clear advantage of one over the other. For example,
while the power form produces the highest percentage of less than 10% of simulation errors, it
has the highest percent of more than 20% simulation errors. On balance, inverse, power, log and
quadratic forms generate similar accuracy in simulation, with two thirds to 3 quarters of
simulations having less than 10% simulation errors, and 12.5% to a quarter of simulations having
more than 20% of simulation errors.

Table 6. Simulations Errors for Alternative Specifications (%)

Top 2…m firm model specification

Simulation Error* Inverse Power Log Exponential Quadratic
<5% 50.0 47.5 32.5 10.0 60.0
5-10% 12.5 25.0 30.0 15.0 5.0
10-15% 12.5 2.5 12.5 0.0 10.0
15-20% 12.5 0.0 10.0 5.0 12.5
>20% 12.5 25.0 15.0 70.0 12.5
* Measured in simulation discrepancy as a percent of the actual HHI

Some additional observations may be made regarding the simulations. An exponential

specification tends to under-estimate the HHIs while a power functional model tends to over-
 31

estimate the HHIs. Many simulations with large distortions seem to occur in 1996, 1997 and
1998. In quite a few specifications, simulations for these several years diverge from the actual
HHIs by over 100 points. A possible reason for such a large difference is that the actual HHI
jumps by 70 points from 1996 to 1997, and over 100 points from 1997 to 1998. Such a large
change in HHI causes significant changes in size distribution. As a result, the parameter
estimates based on the previous year’s data cannot adequately capture the changes, leading to
simulations widely off the mark. However, such reasoning cannot be applied to simulations in
2000 and 2001. The actual HHI decreases by over 100 points from 1999 and 2000, and again
from 2000 to 2001. However, there is no consistent significant error in simulation. The cause for
significant distortions remains an issue for further investigation.

CONCLUDING REMARKS

This paper simulates HHIs by specifying functions of firm size distributions. This is an attempt
to go a step beyond calculating the HHI using data from the same year. The paper proposes three
types of alternative specifications concerning the size distribution of firms in an industry or a
geographic market. The advantage of arithmetic progression and linear function is that HHIs
resultant can be expressed as a function of the size of one (largest) firm. A group of non-linear
specifications where the firm size is a function of firm rank are used in the empirical simulations.
Simulation for the St. Louis banking market is conducted using non-linear specifications, with
various combinations of specifications for dominant firms and remaining firms. Simulation
results show that relatively higher accuracy is achieved for inverse, power, log and quadratic
specifications for the St. Louis bank market between 1994 and 2002.

REFERENCES

Cabral, L.M.B. 2000. Introduction to Industrial Organization. Cambridge, Massachusetts: The

MIT Press

Devine, P.J., Jones, R.M., Lee, N, and Tyson, W.J. 1974. An Introduction to Industrial
Economics. London: George Allen & Unwin Ltd

Herfindahl, O.C. 1959. Copper Costs and Prices: 1870 – 1957. Baltimore: The John Hopkins
Press

Hirschman, A.O. 1964. The Paternity of an Index. 1964. American Economic Review Vol. 54,
pp 761-762.

------ 1980. National Power and the Structure of Foreign Trade. Los Angles: University of
California Press

Jacquemin, A. 1987. The New Industrial Organization: Market forces and strategic behavior.
Cambridge, Massachusetts: the MIT Press

Rosenbluth, G. 1955. Measures of Concentration, In Business Concentration and Price Policy by

National Bureau of Economic Research. Princeton: Princeton University Press
 32

---------. 1957. Concentration in Canadian Manufacturing Industries. Princeton: Princeton

University Press

Shepard, W. G. 1979. The Economics of Industrial Organization. Englewood Cliff, New Jersey:
Prencice-Hall, Inc.

Zhou, B. 1997. Chaning banking market structure amid banking restrcutring: a case of St. Louis.
Papers and Proceedings of the Applied Geography Conferences 20: 110-118