Econometrica, Vol. 70, No.

1 (January, 2002), 191–221

DETERMINING THE NUMBER OF FACTORS IN

APPROXIMATE FACTOR MODELS

By Jushan Bai and Serena Ng1

In this paper we develop some econometric theory for factor models of large dimen-
sions. The focus is the determination of the number of factors
r, which is an unresolved
issue in the rapidly growing literature on multifactor models. We first establish the con-
vergence rate for the factor estimates that will allow for consistent estimation of r. We
then propose some panel criteria and show that the number of factors can be consistently
estimated using the criteria. The theory is developed under the framework of large cross-
sections (N ) and large time dimensions (T ). No restriction is imposed on the relation
between N and T . Simulations show that the proposed criteria have good finite sample
properties in many configurations of the panel data encountered in practice.

Keywords: Factor analysis, asset pricing, principal components, model selection.

1 introduction
The idea that variations in a large number of economic variables can be
modeled by a small number of reference variables is appealing and is used in
many economic analyses. For example, asset returns are often modeled as a
function of a small number of factors. Stock and Watson (1989) used one ref-
erence variable to model the comovements of four main macroeconomic aggre-
gates. Cross-country variations are also found to have common components; see
Gregory and Head (1999) and Forni, Hallin, Lippi, and Reichlin (2000b). More
recently, Stock and Watson (1999) showed that the forecast error of a large num-
ber of macroeconomic variables can be reduced by including diffusion indexes,
or factors, in structural as well as nonstructural forecasting models. In demand
analysis, Engel curves can be expressed in terms of a finite number of factors.
Lewbel (1991) showed that if a demand system has one common factor, budget
shares should be independent of the level of income. In such a case, the num-
ber of factors is an object of economic interest since if more than one factor is
found, homothetic preferences can be rejected. Factor analysis also provides a
convenient way to study the aggregate implications of microeconomic behavior,
as shown in Forni and Lippi (1997).
Central to both the theoretical and the empirical validity of factor models is
the correct specification of the number of factors. To date, this crucial parameter
1
We thank three anonymous referees for their very constructive comments, which led to a much
improved presentation. The first author acknowledges financial support from the National Science
Foundation under Grant SBR-9709508. We would like to thank participants in the econometrics
seminars at Harvard-MIT, Cornell University, the University of Rochester, and the University of
Pennsylvania for help suggestions and comments. Remaining errors are our own.

191
192 jushan bai and serena ng

is often assumed rather than determined by the data.2 This paper develops a
formal statistical procedure that can consistently estimate the number of factors
from observed data. We demonstrate that the penalty for overfitting must be
a function of both N and T (the cross-section dimension and the time dimen-
sion, respectively) in order to consistently estimate the number of factors. Con-
sequently the usual AIC and BIC, which are functions of N or T alone, do
not work when both dimensions of the panel are large. Our theory is developed
under the assumption that both N and T converge to infinity. This flexibility is
of empirical relevance because the time dimension of datasets relevant to factor
analysis, although small relative to the cross-section dimension, is too large to
justify the assumption of a fixed T .
A small number of papers in the literature have also considered the problem
of determining the number of factors, but the present analysis differs from these
works in important ways. Lewbel (1991) and Donald (1997) used the rank of a
matrix to test for the number of factors, but these theories assume either N or T
is fixed. Cragg and Donald (1997) considered the use of information criteria when
the factors are functions of a set of observable explanatory variables, but the data
still have a fixed dimension. For large dimensional panels, Connor and Korajczyk
(1993) developed a test for the number of factors in asset returns, but their test
is derived under sequential limit asymptotics, i.e., N converges to infinity with a
fixed T and then T converges to infinity. Furthermore, because their test is based
on a comparison of variances over different time periods, covariance stationarity
and homoskedasticity are not only technical assumptions, but are crucial for the
validity of their test. Under the assumption that N →  for fixed T , Forni and
Reichlin (1998) suggested a graphical approach to identify√ the number of factors,
but no theory is available. Assuming N  T →  with N /T → , Stock and
Watson (1998) showed that a modification to the BIC can be used to select
the number of factors optimal for forecasting a single series. Their criterion is
restrictive not only because it requires N  T , but also because there can be
factors that are pervasive for a set of data and yet have no predictive ability
for an individual data series. Thus, their rule may not be appropriate outside of
the forecasting framework. Forni, Hallin, Lippi, and Reichlin (2000a) suggested
a multivariate variant of the AIC but neither the theoretical nor the empirical
properties of the criterion are known.
We set up the determination of factors as a model selection problem. In con-
sequence, the proposed criteria depend on the usual trade-off between good fit
and parsimony. However, the problem is nonstandard not only because account
needs to be taken of the sample size in both the cross-section and the time-
series dimensions, but also because the factors are not observed. The theory we
developed does not rely on sequential limits, nor does it impose any restrictions
between N and T . The results hold under heteroskedasticity in both the time and

2
Lehmann and Modest (1988), for example, tested the APT for 5, 10, and 15 factors. Stock and
Watson (1989) assumed there is one factor underlying the coincident index. Ghysels and Ng (1998)
tested the affine term structure model assuming two factors.
 approximate factor models 193

the cross-section dimensions. The results also hold under weak serial and cross-
section dependence. Simulations show that the criteria have good finite sample
properties.
The rest of the paper is organized as follows. Section 2 sets up the preliminaries
and introduces notation and assumptions. Estimation of the factors is considered
in Section 3 and the estimation of the number of factors is studied in Section 4.
Specific criteria are considered in Section 5 and their finite sample properties
are considered in Section 6, along with an empirical application to asset returns.
Concluding remarks are provided in Section 7. All the proofs are given in the
Appendix.

2 factor models
Let Xit be the observed data for the ith cross-section unit at time t, for i =
1     N , and t = 1     T . Consider the following model:

(1) Xit = i Ft + eit 

where Ft is a vector of common factors, i is a vector of factor loadings associated

with Ft , and eit is the idiosyncratic component of Xit . The product i Ft is called
the common component of Xit . Equation (1) is then the factor representation of
the data. Note that the factors, their loadings, as well as the idiosyncratic errors
are not observable.
Factor analysis allows for dimension reduction and is a useful statistical tool.
Many economic analyses fit naturally into the framework given by (1).
1. Arbitrage pricing theory. In the finance literature, the arbitrage pricing the-
ory (APT) of Ross (1976) assumes that a small number of factors can be used to
explain a large number of asset returns. In this case, Xit represents the return of
asset i at time t Ft represents the vector of factor returns, and eit is the idiosyn-
cratic component of returns. Although analytical convenience makes it appealing
to assume one factor, there is growing evidence against the adequacy of a single
factor in explaining asset returns.3 The shifting interest towards use of multifactor
models inevitably calls for a formal procedure to determine the number of fac-
tors. The analysis to follow allows the number of factors to be determined even
when N and T are both large. This is especially suited for financial applications
when data are widely available for a large number of assets over an increasingly
long span. Once the number of factors is determined, the factor returns Ft can
also be consistently estimated (up to an invertible transformation).
2. The rank of a demand system. Let p be a price vector for J goods and ser-
vices, eh be total spending on the J goods by household h. Consumer theory
postulates that Marshallian demand for good j by consumer h is Xjh = gj
p eh .
Let wjh = Xjh /eh be the budget share for household h on the jth good. The
3
Cochrane (1999) stressed that financial economists now recognize that there are multiple sources
of risk, or factors, that give rise to high returns. Backus, Forsei, Mozumdar, and Wu (1997) made
similar conclusions in the context of the market for foreign assets.
194 jushan bai and serena ng

rank of a demand system holding prices fixed is the smallest integer r such that
wj
e = j1 G1
e + · · · + jr Gr
e. Demand systems are of the form (1) where the
r factors, common across goods, are Fh = G1
eh  · · · Gr
eh  . When the number
of households, H , converges to infinity with a fixed J  G1
e · · · Gr
e can be esti-
mated simultaneously, such as by nonparametric methods developed in Donald
(1997). This approach will not work when the number of goods, J , also converges
to infinity. However, the theory to be developed in this paper will still provide a
consistent estimation of r and without the need for nonparametric estimation of
the G
· functions. Once the rank of the demand system is determined, the non-
parametric functions evaluated at eh allow Fh to be consistently estimable (up to
a transformation). Then functions G1
e · · · Gr
e may be recovered (also up to
a matrix transformation) from F
h
h = 1     H  via nonparametric estimation.
3. Forecasting with diffusion indices. Stock and Watson (1998, 1999) considered
forecasting inflation with diffusion indices (“factors”) constructed from a large
number of macroeconomic series. The underlying premise is that these series may
be driven by a small number of unobservable factors. Consider the forecasting
equation for a scalar series

yt+1 =  Ft +  Wt + t 

The variables Wt are observable. Although we do not observe Ft , we observe

Xit  i = 1     N . Suppose Xit bears relation with Ft as in (1). In the present
context, we interpret (1) as the reduced-form representation of Xit in terms of
the unobservable factors. We can first estimate Ft from (1). Denote it by F
t . We
can then regress yt on F
t−1 and Wt−1 to obtain the coefficients ˆ and , ˆ from
which a forecast

ŷT +1T = ˆ  F
T + W
ˆ T

can be formed. Stock and Watson (1998, 1999) showed that this approach of
forecasting outperforms many competing forecasting methods. But as pointed
out earlier, the dimension of F in Stock and Watson (1998, 1999) was determined
using a criterion that minimizes the mean squared forecast errors of y. This may
not be the same as the number of factors underlying Xit , which is the focus of
this paper.

21 Notation and Preliminaries

Let Ft0  0i ,
and r denote the true common factors, the factor loadings, and the
true number of factors, respectively. Note that Ft0 is r dimensional. We assume
that r does not depend on N or T . At a given t, we have
Xt = 0 Ft0 + et
(2)

N × 1
N × r
r × 1
N × 1

where Xt =
X1t  X2t      XNt   0 =
01  02      0N  , and et =
e1t  e2t     

eNt  . Our objective is to determine the true number of factors, r. In classical
 approximate factor models 195

factor analysis (e.g., Anderson (1984)), N is assumed fixed, the factors are inde-
pendent of the errors et , and the covariance of et is diagonal. Normalizing the

covariance matrix of Ft to be an identity matrix, we have = 0 0 + !, where
and ! are the covariance matrices of Xt and et , respectively. Under these
assumptions, a root-T consistent and asymptotically normal estimator of , say,

the sample covariance matrix
=
1/T  Tt=1
Xt − X
X

t − X can be obtained.
The essentials of classical factor analysis carry over to the case of large N but
fixed T since the N × N problem can be turned into a T × T problem, as noted
by Connor and Korajczyk (1993) and others.
Inference on r under classical assumptions can, in theory, be based on the
eigenvalues of
since a characteristic of a panel of data that has an r factor
representation is that the first r largest population eigenvalues of the N × N
covariance of Xt diverge as N increases to infinity, but the
r + 1th eigenvalue
is bounded; see Chamberlain and Rothschild (1983). But it can be shown that all
nonzero sample eigenvalues (not just the first r) of the matrix
increase with
N , and a test based on the sample eigenvalues is thus not feasible. A likelihood
ratio test can also, in theory, be used to select the number of factors if, in addi-
tion, normality of et is assumed. But as found by Dhrymes, Friend, and Glutekin
(1984), the number of statistically significant factors determined by the likelihood
ratio test increases with N even if the true number of factors is fixed. Other
methods have also been developed to estimate the number of factors assuming
the size of one dimension is fixed. But Monte Carlo simulations in Cragg and
Donald (1997) show that these methods tend to perform poorly for moderately
large N and T . The fundamental problem is that the theory developed for clas-
sical factor models does not apply when both N and T → . This is because
consistent estimation of (whether it is an N × N or a T × T matrix) is not a
well defined problem. For example, when N > T , the rank of
is no more than
T , whereas the rank of can always be N . New theories are thus required to
analyze large dimensional factor models.
In this paper, we develop asymptotic results for consistent estimation of the
number of factors when N and T → . Our results complement the sparse but
growing literature on large dimensional factor analysis. Forni and Lippi (2000)
and Forni et al. (2000a) obtained general results for dynamic factor models, while
Stock and Watson (1998) provided some asymptotic results in the context of
forecasting. As in these papers, we allow for cross-section and serial dependence.
In addition, we also allow for heteroskedasticity in et and some weak dependence
between the factors and the errors. These latter generalizations are new in our
analysis. Evidently, our assumptions are more general than those used when the
sample size is fixed in one dimension.
Let X i be a T × 1 vector of time-series observations for the ith cross-section
unit. For a given i, we have

Xi = F 0 0i + ei 
(3)

T × 1
T × r
r × 1
T × 1
196 jushan bai and serena ng

where X i =
Xi1  Xi2      XiT   F 0 =
F10  F20      FT0  , and ei =
ei1  ei2     
eiT  . For the panel of data X =
X 1      X N , we have

X = F0 0 + e
(4)

T × N 
T × r
r × N 
T × N 

with e =
e1      eN .
Let tr
A denote the trace of A. The norm of the matrix A is then A =
tr
A A1/2 . The following assumptions are made:
T 0 0
Assumption A—Factors: EFt0 4 <  and T −1 t=1 Ft Ft → F as T → 
for some positive definite matrix F .

Assumption B—Factor Loadings:  i  ≤ ¯ < , and 0 0 /N − D → 0 as



N →  for some r × r positive definite matrix D.

Assumption C—Time and Cross-Section Dependence and Heteroskedasticity:

There exists a positive constant M < , such that for all N and T ,
1. E
eit  = 0 Eeit 8 ≤ M;

2. E
es et /N  = E
N −1 N i=1 eis eit  = )N
s t )N
s s ≤ M for all s, and
 
T −1 Ts=1 Tt=1 )N
s t ≤ M;
3. E
eit ejt  = *ij t with *ij t  ≤ *ij  for some *ij and for all t; in addition,
 N
N −1 N i=1 j=1 *ij  ≤ M;
 N T T
4. E
eit ejs  = *ij ts and
N T −1 N i=1 j=1 t=1 s=1 *ij ts  ≤ M;

5. for every
t s EN −1/2 N i=1 e is e it − E
e is e it  4
≤ M.

Assumption D—Weak Dependence between Factors and Idiosyncratic Errors:

 N 
 2 
1  1  T 
E  Ft eit 
0
N i=1  √
T t=1  ≤ M

Assumption A is standard for factor models. Assumption B ensures that each

factor has a nontrivial contribution to the variance of Xt . We only consider
nonrandom factor loadings for simplicity. Our results still hold when the i are
random, provided they are independent of the factors and idiosyncratic errors,
and E i 4 ≤ M. Assumption C allows for limited time-series and cross-section
dependence in the idiosyncratic component. Heteroskedasticity in both the time
and cross-section dimensions is also allowed. Under stationarity in the time
dimension, )N
s t = )N
s − t, though the condition is not necessary. Given
Assumption C1, the remaining assumptions in C are easily satisfied if the eit are
independent for all i and t. The allowance for some correlation in the idiosyn-
cratic components sets up the model to have an approximate factor structure. It is
more general than a strict factor model, which assumes eit is uncorrelated across
i, the framework in which the APT theory of Ross (1976) is based. Thus, the
results to be developed will also apply to strict factor models. When the factors
 approximate factor models 197

and idiosyncratic errors are independent (a standard assumption for conventional

factor models), Assumption D is implied by Assumptions A and C. Independence
is not required for D to be true. For example, suppose that eit = it Ft  with
it being independent of Ft and it satisfies Assumption C; then Assumption D
holds. Finally, the developments proceed assuming that the panel is balanced.
We also note that the model being analyzed is static, in the sense that Xit has a
contemporaneous relationship with the factors. The analysis of dynamic models
is beyond the scope of this paper.
For a factor model to be an approximate factor model in the sense of
Chamberlain and Rothschild (1983), the largest eigenvalue (and hence all of
the eigenvalues) of the N × N covariance matrix ! = E
et et  must be bounded.
Note that Chamberlain and Rothschild focused on the cross-section behavior of
the model and did not make explicit assumptions about the time-series behavior
of the model. Our framework allows for serial correlation and heteroskedastic-
ity and is more general than their setup. But if we assume et is stationary with
E
eit ejt  = *ij , then from matrix theory, the largest eigenvalue of ! is bounded
 N
by max i N j=1 *ij . Thus if we assume j=1 *ij  ≤ M for all i and all N , which
implies Assumption C3, then (2) will be an approximate factor model in the
sense of Chamberlain and Rothschild.

3 estimation of the common factors

When N is small, factor models are often expressed in state space form, nor-
mality is assumed, and the parameters are estimated by maximum likelihood. For
example, Stock and Watson (1989) used N = 4 variables to estimate one factor,
the coincident index. The drawback is that because the number of parameters
increases with N ,4 computational difficulties make it necessary to abandon infor-
mation on many series even though they are available.
We estimate common factors in large panels by the method of asymptotic prin-
cipal components.5 The number of factors that can be estimated by this (non-
parametric) method is min+N  T ,, much larger than permitted by estimation
of state space models. But to determine which of these factors are statistically
important, it is necessary to first establish consistency of all the estimated com-
mon factors when both N and T are large. We start with an arbitrary number
k
k < min+N  T ,. The superscript in ki and Ftk signifies the allowance of k
factors in the estimation. Estimates of k and F k are obtained by solving the
optimization problem

N 
T  2
V
k = min
N T −1 Xit − ki Ftk
 F k i=1 t=1

4
Gregory, Head, and Raynauld (1997) estimated a world factor and seven country specific fac-
tors from output, consumption, and investment for each of the G7 countries. The exercise involved
estimation of 92 parameters and perhaps stretched the state-space model to its limit.
5
The method of asymptotic principal components was studied by Connor and Korajzcyk (1986)
and Connor and Korajzcyk (1988) for fixed T . Forni et al. (2000a) and Stock and Watson (1998)
considered the method for large T .
198 jushan bai and serena ng
 
subject to the normalization of either k k /N = Ik or F k F k /T = Ik . If we con-

centrate out k and use the normalization that F k F k /T = Ik , the optimization

to maximizing tr
F k
XX  F k . The estimated factor matrix,
problem is identical √
k
denoted by F , is T times the eigenvectors corresponding to the k largest
k =
F k F k −1 F k X = F k X/T
eigenvalues of the T × T matrix XX  . Given F k  
is the corresponding matrix of factor loadings.
The solution to the above minimization problem is not unique, even though the
sum of squared residuals V
k k k
√ is unique. Another solution is given by
F   ,
k
where  is constructed as N times the eigenvectors corresponding to the k
largest eigenvalues of the N ×N matrix X  X. The normalization that  k 
k /N =
k k
Ik implies F = X  /N . The second set of calculations is computationally less
costly when T > N , while the first is less intensive when T < N .6
Define

F
k = F k
F k F k /T 1/2 


a rescaled estimator of the factors. The following theorem summarizes the asymp-
totic properties of the estimated factors.

Theorem 1: For any fixed k ≥ 1, √ there √exists a
r × k matrix H k with

k
rank
H  = min+k r,, and CN T = min+ N  T ,, such that
 T 
1  
F
k − H k F 0 2 = Op
1
(5) CN2 T
T t=1 t t

Because the true factors (F 0 ) can only be identified up to scale, what is being
considered is a rotation of F 0 . The theorem establishes that the time average of
the squared deviations between the estimated factors and those that lie in the
true factor space vanish as N  T → . The rate of convergence is determined by
the smaller of N or T , and thus depends on the panel structure.
Under the additional assumption that Ts=1 )N
s t2 ≤ M for all t and T , the
result7
 2
CN2 T F
t − H k Ft0  = Op
1 for each t

(6)

can be obtained. Neither Theorem 1 nor (6) implies uniform convergence in t.

Uniform convergence is considered by Stock and Watson (1998). These authors
obtained a much slower convergence rate than CN2 T , and their result requires
√
N  T . An important insight of this paper is that, to consistently estimate
the number of factors, neither (6) nor uniform convergence is required. It is
the average convergence rate of Theorem 1 that is essential. However, (6) could
be useful for statistical analysis on the estimated factors and is thus a result of
independent interest.
6
A more detailed account of computation issues, including how to deal with unbalanced panels, is
given in Stock and Watson (1998).
7
The proof is actually simpler than that of Theorem 1 and is thus omitted to avoid repetition.
 approximate factor models 199

4 estimating the number of factors

Suppose for the moment that we observe all potentially informative factors
but not the factor loadings. Then the problem is simply to choose k factors that
best capture the variations in X and estimate the corresponding factor loadings.
Since the model is linear and the factors are observed, i can be estimated by
applying ordinary least squares to each equation. This is then a classical model
selection problem. A model with k + 1 factors can fit no worse than a model
with k factors, but efficiency is lost as more factor loadings are being estimated.
Let F k be a matrix of k factors, and
1  N  T 
 2
V
k F k  = min Xit − ki Ftk
 N T i=1 t=1

be the sum of squared residuals (divided by N T ) from time-series regressions of

X i on the k factors for all i. Then a loss function V
k F k  + kg
N  T , where
g
N  T  is the penalty for overfitting, can be used to determine k. Because the
estimation of i is classical, it can be shown that the BIC with g
N  T  = ln
T /T
can consistently estimate r. On the other hand, the AIC with g
N  T  = 2/T
may choose k > r even in large samples. The result is the same as in Geweke
and Meese (1981) derived for N = 1 because when the factors are observed,
the penalty factor does not need to take into account the sample size in the
cross-section dimension. Our main result is to show that this will no longer be
true when the factors have to be estimated, and even the BIC will not always
consistently estimate r.
Without loss of generality, we let

1 N  T  2
V
k F
k  = min Xit − ki F
tk

(7)
 NT
i=1 t=1

denote the sum of squared residuals (divided by N T ) when k factors are esti-
mated. This sum of squared residuals does not depend on which estimate of F is
used because they span the same vector space. That is, V
k F k  = V
k F k  =
V
k F
k . We want to find penalty functions, g
N  T , such that criteria of the
form

PC
k = V
k F
k  + kg
N  T 

can consistently estimate r. Let kmax be a bounded integer such that r ≤ kmax.

Theorem 2: Suppose that Assumptions A–D hold and that the k factors
are estimated by principal components. Let k̂ = arg min0≤k≤kmax PC
k. Then
2
limN  T → Probk̂ = r = 1 if√(i) g
N
√  T  → 0 and (ii) CN T · g
N  T  →  as N ,
T → , where CN T = min+ N  T ,.

Conditions (i) and (ii) are necessary in the sense that if one of the conditions
is violated, then there will exist a factor model satisfying Assumptions A–D, and
200 jushan bai and serena ng

yet the number of factors cannot be consistently estimated. However, conditions

(i) and (ii) are not always required to obtain a consistent estimate of r.
A formal proof of Theorem 2 is provided in the Appendix. The crucial ele-
ment in consistent estimation of r is a penalty factor that vanishes at an appro-
priate rate such that under and overparameterized models will not be chosen.
An implication of Theorem 2 is the following:

Corollary 1: Under the Assumptions of Theorem 2, the class of criteria

defined by

IC
k = ln
V
k F
k  + kg
N  T 

will also consistently estimate r.

Note that V
k F
k  is simply the average residual variance when k factors are
assumed for each cross-section unit. The IC criteria thus resemble information
criteria frequently used in time-series analysis, with the important difference that
the penalty here depends on both N and T .
Thus far, it has been assumed that the common factors are estimated by the
method of principle components. Forni and Reichlin (1998) and Forni et al.
(2000a) studied alternative estimation methods. However the proof of Theorem 2
mainly uses the fact that F
t satisfies Theorem 1, and does not rely on principal
components per se. We have the following corollary:

Corollary 2: Let G
k be an arbitrary estimator of F 0 . Suppose there exists a

k k  = min+k r,, and for some C
matrix H such that rank
H 2 ≤ C 2 ,
NT NT

 T  
1 
(8) 2
C G k F 0 2 = Op
1

k − H
NT t t
T t=1

k and CN T replaced by C
Then Theorem 2 still holds with F
k replaced by G N T .

The sequence of constants C 2 does not need to equal C 2 = min+N  T ,.

NT NT

t satisfying
Theorem 2 holds for any estimation method that yields estimators G
8 2
(8). Naturally, the penalty would then depend on CN T , the convergence rate
for G
t .

5 the PCp and the ICp

In this section, we assume that the method of principal components is used to

estimate the factors and propose specific formulations of g
N  T  to be used in

8
We are grateful for a referee whose question led to the results reported here.
 approximate factor models 201
N T
practice. Let 3̂ 2 be a consistent estimate of
N T −1 i=1 t=1 E
eit 
2
. Consider
the following criteria:
   
N +T NT
PCp1
k = V
k F
k  + k3̂ 2 ln 4
NT N +T
 

k 2 N +T
PCp2
k = V
k F  + k3̂ ln CN2 T 4
NT
 2


k 2 ln CN T
PCp3
k = V
k F  + k3̂ 
CN2 T

Since V
k F
k  = N −1 N 2 2 
i=1 3̂i , where 3̂i = êi êi /T , the criteria generalize the Cp
criterion of Mallows (1973) developed for selection of models in strict time-
series or cross-section contexts to a panel data setting. For this reason, we refer
to these statistics as Panel Cp
PCp  criteria. Like the Cp criterion, 3̂ 2 provides
the proper scaling to the penalty term. In applications, it can be replaced by
V
kmax F
kmax . The proposed penalty functions are based on the sample size in
the smaller of the two dimensions. All three criteria satisfy conditions (i) and (ii)
of Theorem 2 since CN−2T ≈

N + T /N T  → 0 as N  T → . However, in finite
samples, CN−2T ≤
N + T /N T . Hence, the three criteria, although asymptotically
equivalent, will have different properties in finite samples.9
Corollary 1 leads to consideration of the following three criteria:
   

k N +T NT
ICp1
k = ln
V
k F  + k ln 4
NT N +T
 
N +T
(9) ICp2
k = ln
V
k F
k  + k ln CN2 T 4
NT
 
ln CN2 T
ICp3
k = ln
V
k F
k  + k 
CN2 T

The main advantage of these three panel information criteria (ICp ) is that they
do not depend on the choice of kmax through 3̂ 2 , which could be desirable in
practice. The scaling by 3̂ 2 is implicitly performed by the logarithmic transfor-
mation of V
k F
k  and thus not required in the penalty term.
The proposed criteria differ from the conventional Cp and information criteria
used in time-series analysis in that g
N  T  is a function of both N and T . To
understand why the penalty must be specified as a function of the sample size in

9
Note that PCp1 and PCp2 , and likewise, ICp1 and ICp2 , apply specifically to the principal com-
ponents estimator because CN2 T = min+N  T , is used in deriving them. For alternative estimators
N T .
satisfying Corollary 2, criteria PCp3 and ICp3 are still applicable with CN T replaced by C
202 jushan bai and serena ng

both dimensions, consider the following:

 

k 2 2
AIC1
k = V
k F  + k3̂ 4
T
 
ln T
BIC1
k = V
k F
k  + k3̂ 2 4
T
 

k 2 2
AIC2
k = V
k F  + k3̂ 4
N
 
ln N
BIC2
k = V
k F
k  + k3̂ 2 4
N
 

k 2
N + T − k
AIC3
k = V
k F  + k3̂ 2 4
NT
 

N + T − k ln
N T 
BIC3
k = V
k F
k  + k3̂ 2 
NT
The penalty factors in AIC1 and BIC1 are standard in time-series applications.
Although g
N  T  → 0 as T →  AIC1 fails the second condition of Theorem 2
for all N and T . When N  T and N log
T /T −→ , the BIC1 also fails
condition (ii) of Theorem 2. Thus we expect the AIC1 will not work for all N
and T , while the BIC1 will not work for small N relative to T . By analogy, AIC2
also fails the conditions of Theorem 2, while BIC2 will work only if N  T .
The next two criteria, AIC3 and BIC3 , take into account the panel nature of the
problem. The two specifications of g
N  T  reflect first, that the effective number
of observations is N · T , and second, that the total number of parameters being
estimated is k
N + T − k. It is easy to see that AIC3 fails the second condition
of Theorem 2. While the BIC3 satisfies this condition, g
N  T  does not always
vanish. For example, if N = exp
T , then g
N  T  → 1 and the first condition
of Theorem 2 will not be satisfied. Similarly, g
N  T  does not vanish when T =
exp
N . Therefore BIC3 may perform well for some but not all configurations
of the data. In contrast, the proposed criteria satisfy both conditions stated in
Theorem 2.

6 simulations and an empirical application

We first simulate data from the following model:

r √
Xit = ij Ftj + 6eit
j=1
√
= cit + 6eit 

where the factors are T × r matrices of N
0 1 variables, and the factor load-
ings are N
0 1 variates. Hence, the common component of Xit , denoted by cit ,
has variance r. Results with ij uniformly distributed are similar and will not
 approximate factor models 203

be reported. Our base case assumes that the idiosyncratic component has the
same variance as the common component (i.e. 6 = r. We consider thirty con-
figurations of the data. The first five simulate plausible asset pricing applications
with five years of monthly data (T = 60) on 100 to 2000 asset returns. We then
increase T to 100. Configurations with N = 60 T = 100 and 200 are plausible
sizes of datasets for sectors, states, regions, and countries. Other configurations
are considered to assess the general properties of the proposed criteria. All com-
putations were performed using Matlab Version 5.3.
Reported in Tables I to III are the averages of k̂ over 1000 replications, for
r = 1 3, and 5 respectively, assuming that eit is homoskedastic N
0 1. For all
cases, the maximum number of factors, kmax, is set to 8.10 Prior to computation
of the eigenvectors, each series is demeaned and standardized to have unit vari-
ance. Of the three PCp criteria that satisfy Theorem 2, PCp3 is less robust than
PCp1 and PCp2 when N or T is small. The ICp criteria generally have prop-
erties very similar to the PCp criteria. The term N T /
N + T  provides a small
sample correction to the asymptotic convergence rate of CN2 T and has the effect
of adjusting the penalty upwards. The simulations show this adjustment to be
desirable. When min+N  T , is 40 or larger, the proposed tests give precise esti-
mates of the number of factors. Since our theory is based on large N and T , it is
not surprising that for very small N or T , the proposed criteria are inadequate.
Results reported in the last five rows of each table indicate that the ICp crite-
ria tend to underparameterize, while the PCp tend to overparameterize, but the
problem is still less severe than the AIC and the BIC, which we now consider.
The AIC and BIC’s that are functions of only N or T have the tendency
to choose too many factors. The AIC3 performs somewhat better than AIC1
and AIC2 , but still tends to overparameterize. At first glance, the BIC3 appears
to perform well. Although BIC3 resembles PCp2 , the former penalizes an extra
factor more heavily since ln
N T  > ln CN2 T . As can be seen from Tables II and
III, the BIC3 tends to underestimate r, and the problem becomes more severe
as r increases.
Table IV relaxes the assumption of homoskedasticity. Instead, we let eit = eit1
for t odd, and eit = eit1 + eit2 for t even, where eit1 and eit2 are independent N
0 1.
Thus, the variance in the even periods is twice as large as the odd periods.
Without loss of generality, we only report results for r = 5. PCp1  PCp2  ICp1 , and
ICp2 continue to select the true number of factors very accurately and dominate
the remaining criteria considered.
We then vary the variance of the idiosyncratic errors relative to the common
component. When 6 < r, the variance of the common component is relatively
large. Not surprisingly, the proposed criteria give precise estimates of r. The
results will not be reported without loss of generality. Table V considers the case
6 = 2r. Since the variance of the idiosyncratic component is larger than the

10
In time-series analysis, a rule such as 8 int[
T /1001/4 ] considered in Schwert (1989) is sometimes
used to set kmax, but no such guide is available for panel analysis. Until further results are available,
a rule that replaces T in Schwert’s rule by min+N  T , could be considered.
204 jushan bai and serena ng

TABLE I
r √
DGP: Xit = j=1 ij Ftj + 6eit ; r = 1; 6 = 1.

N T PCp1 PCp2 PCp3 ICp1 ICp2 ICp3 AIC1 BIC1 AIC2 BIC2 AIC3 BIC3

100 40 102 100 297 100 100 100 800 297 800 800 757 100
100 60 100 100 241 100 100 100 800 241 800 800 711 100
200 60 100 100 100 100 100 100 800 100 800 800 551 100
500 60 100 100 100 100 100 100 521 100 800 800 157 100
1000 60 100 100 100 100 100 100 100 100 800 800 100 100
2000 60 100 100 100 100 100 100 100 100 800 800 100 100
100 100 100 100 324 100 100 100 800 324 800 324 668 100
200 100 100 100 100 100 100 100 800 100 800 800 543 100
500 100 100 100 100 100 100 100 800 100 800 800 155 100
1000 100 100 100 100 100 100 100 108 100 800 800 100 100
2000 100 100 100 100 100 100 100 100 100 800 800 100 100
40 100 101 100 269 100 100 100 800 800 800 269 733 100
60 100 100 100 225 100 100 100 800 800 800 225 699 100
60 200 100 100 100 100 100 100 800 800 800 100 514 100
60 500 100 100 100 100 100 100 800 800 467 100 132 100
60 1000 100 100 100 100 100 100 800 800 100 100 100 100
60 2000 100 100 100 100 100 100 800 800 100 100 100 100
4000 60 100 100 100 100 100 100 100 100 800 800 100 100
4000 100 100 100 100 100 100 100 100 100 800 800 100 100
8000 60 100 100 100 100 100 100 100 100 800 800 100 100
8000 100 100 100 100 100 100 100 100 100 800 800 100 100
60 4000 100 100 100 100 100 100 800 800 100 100 100 100
100 4000 100 100 100 100 100 100 800 800 100 100 100 100
60 8000 100 100 100 100 100 100 800 800 100 100 100 100
100 8000 100 100 100 100 100 100 800 800 100 100 100 100
10 50 800 800 800 800 800 800 800 800 800 800 800 718
10 100 800 800 800 800 800 800 800 800 800 800 800 588
20 100 473 394 629 100 100 100 800 800 800 629 800 100
100 10 800 800 800 800 800 800 800 800 800 800 800 800
100 20 562 481 716 100 100 100 800 716 800 800 800 100
Notes: Table I–Table VIII report the estimated number of factors (k̂) averaged over 1000 simulations. The true number of factors
is r and kmax = 8. When the average of k̂ is an integer, the corresponding standard error is zero. In the few cases when the averaged
k̂ over replications is not an integer, the standard errors are no larger than .6. In view of the precision of the estimates in the majority
of cases, the standard errors in the simulations are not reported. The last five rows of each table are for models of small dimensions
(either N or T is small).

common component, one might expect the common factors to be estimated with
less precision. Indeed, ICp1 and ICp2 underestimate r when min+N  T , < 60, but
the criteria still select values of k that are very close to r for other configurations
of the data.
The models considered thus far have idiosyncratic errors that are uncorrelated
across units and across time. For these strict factor models, the preferred criteria
are PCp1  PCp2  IC1 , and IC2 . It should be emphasized that the results reported
are the averages of k̂ over 1000 simulations. We do not report the standard
deviations of these averages because they are identically zero except for a few
 approximate factor models 205

TABLE II
r √
DGP: Xit = j=1 ij Ftj + 6eit ; r = 3; 6 = 3.

N T PCp1 PCp2 PCp3 ICp1 ICp2 ICp3 AIC1 BIC1 AIC2 BIC2 AIC3 BIC3

100 40 300 300 390 300 300 300 800 390 800 800 782 290
100 60 300 300 354 300 300 300 800 354 800 800 753 298
200 60 300 300 300 300 300 300 800 300 800 800 614 300
500 60 300 300 300 300 300 300 595 300 800 800 313 300
1000 60 300 300 300 300 300 300 300 300 800 800 300 300
2000 60 300 300 300 300 300 300 300 300 800 800 300 300
100 100 300 300 423 300 300 300 800 423 800 423 720 300
200 100 300 300 300 300 300 300 800 300 800 800 621 300
500 100 300 300 300 300 300 300 800 300 800 800 315 300
1000 100 300 300 300 300 300 300 301 300 800 800 300 300
2000 100 300 300 300 300 300 300 300 300 800 800 300 300
40 100 300 300 370 300 300 300 800 800 800 370 763 292
60 100 300 300 342 300 300 300 800 800 800 342 739 299
60 200 300 300 300 300 300 300 800 800 800 300 583 300
60 500 300 300 300 300 300 300 800 800 544 300 303 300
60 1000 300 300 300 300 300 300 800 800 300 300 300 300
60 2000 300 300 300 300 300 300 800 800 300 300 300 300
4000 60 300 300 300 300 300 300 300 300 800 800 300 298
4000 100 300 300 300 300 300 300 300 300 800 800 300 300
8000 60 300 300 300 300 300 300 300 300 800 800 300 297
8000 100 300 300 300 300 300 300 300 300 800 800 300 300
60 4000 300 300 300 300 300 300 800 800 300 300 300 299
100 4000 300 300 300 300 300 300 800 800 300 300 300 300
60 8000 300 300 300 300 300 300 800 800 300 300 300 298
100 8000 300 300 300 300 300 300 800 800 300 300 300 300
10 50 800 800 800 800 800 800 800 800 800 800 800 721
10 100 800 800 800 800 800 800 800 800 800 800 800 601
20 100 522 457 662 295 292 298 800 800 800 662 800 268
100 10 800 800 800 800 800 800 800 800 800 800 800 800
100 20 600 529 739 295 291 299 800 739 800 800 800 272

cases for which the average itself is not an integer. Even for these latter cases,
the standard deviations do not exceed 0.6.
We next modify the assumption on the idiosyncratic errors to allow for serial
and cross-section correlation. These errors are generated from the process

J
eit = 8eit−1 + vit + vi−jt 
j=0 j=−J

The case of pure serial correlation obtains when the cross-section correlation
parameter  is zero. Since for each i, the unconditional variance of eit is
1/
1 − 82 , the more persistent are the idiosyncratic errors, the larger are their
variances relative to the common factors, and the precision of the estimates can
be expected to fall. However, even with 8 = 5, Table VI shows that the estimates
provided by the proposed criteria are still very good. The case of pure cross-
206 jushan bai and serena ng

TABLE III
r √
DGP: Xit = j=1 ij Ftj + 6eit ; r = 5; 6 = 5.

N T PCp1 PCp2 PCp3 ICp1 ICp2 ICp3 AIC1 BIC1 AIC2 BIC2 AIC3 BIC3

100 40 499 498 517 488 468 499 800 517 800 800 794 305
100 60 500 500 507 499 494 500 800 507 800 800 787 350
200 60 500 500 500 500 500 500 800 500 800 800 691 380
500 60 500 500 500 500 500 500 688 500 800 800 501 388
1000 60 500 500 500 500 500 500 500 500 800 800 500 382
2000 60 500 500 500 500 500 500 500 500 800 800 500 359
100 100 500 500 542 500 500 501 800 542 800 542 775 416
200 100 500 500 500 500 500 500 800 500 800 800 706 480
500 100 500 500 500 500 500 500 800 500 800 800 502 497
1000 100 500 500 500 500 500 500 500 500 800 800 500 498
2000 100 500 500 500 500 500 500 500 500 800 800 500 498
40 100 500 499 509 486 469 500 800 800 800 509 786 296
60 100 500 500 505 499 494 500 800 800 800 505 781 346
60 200 500 500 500 500 500 500 800 800 800 500 671 383
60 500 500 500 500 500 500 500 800 800 644 500 500 391
60 1000 500 500 500 500 500 500 800 800 500 500 500 379
60 2000 500 500 500 500 500 500 800 800 500 500 500 358
4000 60 500 500 500 500 500 500 500 500 800 800 500 337
4000 100 500 500 500 500 500 500 500 500 800 800 500 496
8000 60 500 500 500 500 500 500 500 500 800 800 500 310
8000 100 500 500 500 500 500 500 500 500 800 800 500 493
60 4000 500 500 500 500 500 500 800 800 500 500 500 335
100 4000 500 500 500 500 500 500 800 800 500 500 500 496
60 8000 500 500 500 500 500 500 800 800 500 500 500 312
100 8000 500 500 500 500 500 500 800 800 500 500 500 493
10 50 800 800 800 800 800 800 800 800 800 800 800 728
10 100 800 800 800 800 800 800 800 800 800 800 800 630
20 100 588 541 699 417 379 468 800 800 800 699 800 279
100 10 800 800 800 800 800 800 800 800 800 800 800 800
100 20 649 594 762 424 387 481 800 762 800 800 800 293

section dependence obtains with 8 = 0. As in Chamberlain and Rothschild (1983),

our theory permits some degree of cross-section correlation. Given the assumed
process for eit , the amount of cross correlation depends on the number of units
that are cross correlated (2J ), as well as the magnitude of the pairwise correla-
tion (). We set  to .2 and J to max+N /20 10,. Effectively, when N
200, 10
percent of the units are cross correlated, and when N > 200 20/N of the sam-
ple is cross correlated. As the results in Table VII indicate, the proposed criteria
still give very good estimates of r and continue to do so for small variations in
 and J . Table VIII reports results that allow for both serial and cross-section
correlation. The variance of the idiosyncratic errors is now
1 + 2J2 /
1 − 82 
times larger than the variance of the common component. While this reduces the
precision of the estimates somewhat, the results generally confirm that a small
degree of correlation in the idiosyncratic errors will not affect the properties of
 approximate factor models 207

TABLE IV
r √ 1 2
DGP: Xit = F
j=1 ij tj + 6e it ; eit = e it + :t it (:t = 1 for t Even, :t = 0 for t Odd); r = 5; 6 = 5.
e

N T PCp1 PCp2 PCp3 ICp1 ICp2 ICp3 AIC1 BIC1 AIC2 BIC2 AIC3 BIC3

100 40 496 486 609 409 337 493 800 609 800 800 800 181
100 60 499 490 585 469 418 501 800 585 800 800 800 208
200 60 500 499 500 493 487 500 800 500 800 800 800 222
500 60 500 500 500 499 498 500 800 500 800 800 791 223
1000 60 500 500 500 500 500 500 797 500 800 800 647 202
2000 60 500 500 500 500 500 500 551 500 800 800 503 172
100 100 500 498 660 498 479 524 800 660 800 660 800 256
200 100 500 500 500 500 500 500 800 500 800 800 800 333
500 100 500 500 500 500 500 500 800 500 800 800 794 393
1000 100 500 500 500 500 500 500 800 500 800 800 613 398
2000 100 500 500 500 500 500 500 536 500 800 800 500 385
40 100 494 480 539 404 330 490 800 800 800 539 799 168
60 100 498 488 541 466 414 500 800 800 800 541 799 204
60 200 500 499 500 495 487 500 800 800 800 500 756 214
60 500 500 500 500 499 498 500 800 800 729 500 507 213
60 1000 500 500 500 500 500 500 800 800 500 500 500 190
60 2000 500 500 500 500 500 500 800 800 500 500 500 159
4000 60 500 500 500 500 500 500 500 500 800 800 500 146
4000 100 500 500 500 500 500 500 500 500 800 800 500 367
8000 60 500 500 500 500 500 500 500 500 800 800 500 116
8000 100 500 500 500 500 500 500 500 500 800 800 500 337
60 4000 500 500 500 500 500 500 800 800 500 500 500 130
100 4000 500 500 500 500 500 500 800 800 500 500 500 362
60 8000 500 500 500 500 500 500 800 800 500 500 500 108
100 8000 500 500 500 500 500 500 800 800 500 500 500 329
10 50 800 800 800 800 800 800 800 800 800 800 800 727
10 100 800 800 800 800 800 800 800 800 800 800 800 634
20 100 613 562 723 285 223 393 800 800 800 723 800 186
100 10 800 800 800 800 800 800 800 800 800 800 800 800
100 20 752 699 799 331 264 617 800 799 800 800 800 230

the estimates. However, it will generally be true that for the proposed criteria to
be as precise in approximate as in strict factor models, N has to be fairly large
relative to J   cannot be too large, and the errors cannot be too persistent as
required by theory. It is also noteworthy that the BIC3 has very good properties
in the presence of cross-section correlations (see Tables VII and VIII) and the
criterion can be useful in practice even though it does not satisfy all the condi-
tions of Theorem 2.

61 Application to Asset Returns

Factor models for asset returns are extensively studied in the finance litera-
ture. An excellent summary on multifactor asset pricing models can be found in
Campbell, Lo, and Mackinlay (1997). Two basic approaches are employed. One
208 jushan bai and serena ng

TABLE V
r √
DGP: Xit = j=1 ij Ftj +
6eit ; r = 5; 6 = r × 2.

N T PCp1 PCp2 PCp3 ICp1 ICp2 ICp3 AIC1 BIC1 AIC2 BIC2 AIC3 BIC3

100 40 463 429 514 279 191 447 800 800 800 514 793 082
100 60 478 441 506 373 261 496 800 800 800 506 786 092
200 60 490 480 500 442 403 494 800 800 800 500 692 093
500 60 496 494 499 477 468 492 800 800 688 499 501 077
1000 60 497 497 498 488 486 493 800 800 500 498 500 056
2000 60 498 498 499 491 489 492 800 800 500 499 500 034
100 100 496 467 542 464 361 501 800 542 800 542 774 123
200 100 500 499 500 498 490 500 800 800 800 500 705 180
500 100 500 500 500 500 500 500 800 800 800 500 502 219
1000 100 500 500 500 500 500 500 800 800 500 500 500 217
2000 100 500 500 500 500 500 500 800 800 500 500 500 206
40 100 461 425 507 265 184 448 800 507 800 800 783 074
60 100 476 438 505 366 260 497 800 505 800 800 781 092
60 200 490 478 500 443 407 495 800 500 800 800 670 088
60 500 497 495 499 478 471 493 644 499 800 800 500 074
60 1000 498 497 499 487 484 492 500 499 800 800 500 051
60 2000 499 498 499 489 488 492 500 499 800 800 500 032
4000 60 499 499 499 492 492 493 800 800 500 499 500 018
4000 100 500 500 500 500 500 500 800 800 500 500 500 172
8000 60 499 499 499 492 492 493 800 800 500 499 500 008
8000 100 500 500 500 500 500 500 800 800 500 500 500 140
60 4000 499 499 499 493 492 495 500 499 800 800 500 015
100 4000 500 500 500 500 500 500 500 500 800 800 500 170
60 8000 499 499 499 492 492 493 500 499 800 800 500 008
100 8000 500 500 500 500 500 500 500 500 800 800 500 140
100 10 800 800 800 800 800 800 800 800 800 800 800 724
100 20 800 800 800 800 800 800 800 800 800 800 800 618
10 50 573 522 690 167 133 279 800 690 800 800 800 112
10 100 800 800 800 800 800 800 800 800 800 800 800 800
20 100 639 579 757 185 144 304 800 800 800 757 800 131

is statistical factor analysis of unobservable factors, and the other is regression

analysis on observable factors. For the first approach, most studies use grouped
data (portfolios) in order to satisfy the small N restriction imposed by classical
factor analysis, with exceptions such as Connor and Korajczyk (1993). The second
approach uses macroeconomic and financial market variables that are thought
to capture systematic risks as observable factors. With the method developed in
this paper, we can estimate the number of factors for the broad U.S. stock mar-
ket, without the need to group the data, or without being specific about which
observed series are good proxies for systematic risks.
Monthly data between 1994.1–1998.12 are available for the returns of 8436
stocks traded on the New York Stock Exchange, AMEX, and NASDAQ. The
data include all lived stocks on the last trading day of 1998 and are obtained
from the CRSP data base. Of these, returns for 4883 firms are available for each
 approximate factor models 209

TABLE VI
r √ 
DGP: Xit = F
j=1 ij tj + 6e it ; e it = 8e it−1 + vit + Jj=−J  j=0 vi−jt ; r = 5; 6 = 5, 8 = 5,  = 0, J = 0.

N T PCp1 PCp2 PCp3 ICp1 ICp2 ICp3 AIC1 BIC1 AIC2 BIC2 AIC3 BIC3

100 40 731 659 800 552 453 800 800 800 800 800 800 297
100 60 611 527 800 500 476 800 800 800 800 800 800 309
200 60 594 538 788 501 499 739 800 788 800 800 800 331
500 60 568 539 679 500 500 511 800 679 800 800 800 341
1000 60 541 527 602 500 500 500 800 602 800 800 800 327
2000 60 521 514 550 500 500 500 800 550 800 800 800 306
100 100 504 500 800 500 497 800 800 800 800 800 800 345
200 100 500 500 775 500 500 712 800 775 800 800 800 426
500 100 500 500 521 500 500 500 800 521 800 800 800 468
1000 100 500 500 500 500 500 500 800 500 800 800 800 473
2000 100 500 500 500 500 500 500 800 500 800 800 800 469
40 100 537 505 730 458 408 582 800 800 800 730 800 245
60 100 513 499 788 493 467 740 800 800 800 788 800 280
60 200 500 500 502 499 496 500 800 800 800 502 800 284
60 500 500 500 500 500 500 500 800 800 800 500 753 272
60 1000 500 500 500 500 500 500 800 800 572 500 504 254
60 2000 500 500 500 500 500 500 800 800 500 500 500 228
4000 60 511 508 522 500 500 500 800 522 800 800 800 281
4000 100 500 500 500 500 500 500 800 500 800 800 800 462
8000 60 505 505 508 500 500 500 800 508 800 800 800 255
8000 100 500 500 500 500 500 500 800 500 800 800 800 437
60 4000 500 500 500 500 500 500 800 800 500 500 500 192
100 4000 500 500 500 500 500 500 800 800 500 500 500 421
60 8000 500 500 500 500 500 500 800 800 500 500 500 164
100 8000 500 500 500 500 500 500 800 800 500 500 500 397
100 10 800 800 800 800 800 800 800 800 800 800 800 747
100 20 800 800 800 800 800 800 800 800 800 800 800 669
10 50 716 668 789 357 292 570 800 800 800 789 800 242
10 100 800 800 800 800 800 800 800 800 800 800 800 800
20 100 800 799 800 793 758 800 800 800 800 800 800 392

of the 60 months. We use the proposed criteria to determine the number of

factors. We transform the data so that each series is mean zero. For this balanced
panel with T = 60 N = 4883 and kmax = 15, the recommended criteria, namely,
PCp1  PCp2  ICp1 , and ICp2 , all suggest the presence of two factors.

7 concluding remarks
In this paper, we propose criteria for the selection of factors in large dimen-
sional panels. The main appeal of our results is that they are developed under
the assumption that N  T →  and are thus appropriate for many datasets typi-
cally used in macroeconomic analysis. Some degree of correlation in the errors is
also allowed. The criteria should be useful in applications in which the number
of factors has traditionally been assumed rather than determined by the data.
210 jushan bai and serena ng

TABLE VII
r √ J
DGP: Xit = F
j=1 ij tj + 6e it ; e it = 8e it−1 + vit + j=−J  j=0 vi−jt ; r = 5; 6 = 5, 8 = 00,  = 20,
J = max+N /20 10,.

N T PCp1 PCp2 PCp3 ICp1 ICp2 ICp3 AIC1 BIC1 AIC2 BIC2 AIC3 BIC3

100 40 550 527 602 509 501 563 800 602 800 800 798 424
100 60 557 524 603 515 502 596 800 603 800 800 796 472
200 60 597 594 600 588 576 599 800 600 800 800 763 489
500 60 501 501 510 500 500 501 744 510 800 800 600 493
1000 60 500 500 500 500 500 500 598 500 800 800 593 493
2000 60 500 500 500 500 500 500 505 500 800 800 501 488
100 100 579 530 631 543 504 603 800 631 800 631 795 498
200 100 600 600 600 600 598 600 800 600 800 800 784 500
500 100 521 511 564 506 503 541 800 564 800 800 602 500
1000 100 500 500 500 500 500 500 600 500 800 800 600 500
2000 100 500 500 500 500 500 500 572 500 800 800 541 500
40 100 517 506 595 500 498 530 800 800 800 595 796 422
60 100 530 506 601 503 500 587 800 800 800 601 794 469
60 200 535 516 595 504 501 565 800 800 800 595 739 489
60 500 543 529 583 505 502 535 800 800 749 583 604 494
60 1000 555 545 579 508 505 525 800 800 601 579 600 493
60 2000 564 559 576 507 504 517 800 800 600 576 600 491
4000 60 500 500 500 500 500 500 500 500 800 800 500 484
4000 100 500 500 500 500 500 500 500 500 800 800 500 500
8000 60 500 500 500 500 500 500 500 500 800 800 500 472
8000 100 500 500 500 500 500 500 500 500 800 800 500 500
60 4000 565 563 572 505 504 509 800 800 600 572 600 485
100 4000 600 600 600 600 600 600 800 800 614 600 602 500
60 8000 567 566 571 504 504 505 800 800 600 571 600 477
100 8000 600 600 600 600 600 600 800 800 600 600 600 500
100 10 800 800 800 800 800 800 800 800 800 800 800 734
100 20 800 800 800 800 800 800 800 800 800 800 800 649
10 50 623 584 718 482 467 514 800 800 800 718 800 372
10 100 800 800 800 800 800 800 800 800 800 800 800 800
20 100 675 627 775 497 473 571 800 775 800 800 800 381

Our discussion has focused on balanced panels. However, as discussed in Rubin

and Thayer (1982) and Stock and Watson (1998), an iterative EM algorithm
can be used to handle missing data. The idea is to replace Xit by its value as
predicted by the parameters obtained from the last iteration when Xit is not
observed. Thus, if i
j and Ft
j are estimated values of i and Ft from the
jth iteration, let Xit∗
j − 1 = Xit if Xit is observed, and Xit∗
j − 1 = i
j − 1×
Ft
j − 1 otherwise. We then minimize V ∗
k with respect to F
j and 
j,
 
where V ∗
k =
N T −1 Ti=1 Tt=1
Xit∗
j − 1 − ki
jFtk
j2 . Essentially, eigen-
values are computed for the T × T matrix X ∗
j − 1X ∗
j − 1 . This process is
iterated until convergence is achieved.
Many issues in factor analysis await further research. Except for some results
derived for classical factor models, little is known about the limiting distribution
 approximate factor models 211

TABLE VIII
r √ J
DGP: Xit = F
j=1 ij tj + 6e it ; e it = 8e it−1 + vit + j=−J  j=0 vi−jt ; r = 5; 6 = 5, 8 = 050,  = 20,
J = max+N /20 10,.

N T PCp1 PCp2 PCp3 ICp1 ICp2 ICp3 AIC1 BIC1 AIC2 BIC2 AIC3 BIC3

100 40 754 692 800 643 552 800 800 800 800 800 800 414
100 60 657 593 800 568 528 800 800 800 800 800 800 439
200 60 652 615 797 600 591 784 800 797 800 800 800 468
500 60 616 597 712 540 530 592 800 712 800 800 800 476
1000 60 571 556 620 503 502 508 800 620 800 800 800 476
2000 60 533 526 561 500 500 500 800 561 800 800 800 469
100 100 598 571 800 572 527 800 800 800 800 800 800 480
200 100 601 600 795 600 599 778 800 795 800 800 800 503
500 100 589 581 606 559 546 594 800 606 800 800 800 500
1000 100 513 509 537 501 501 509 800 537 800 800 800 500
2000 100 500 500 500 500 500 500 800 500 800 800 800 500
40 100 588 546 755 507 493 657 800 800 800 755 800 376
60 100 584 545 796 524 505 779 800 800 800 796 800 425
60 200 567 544 599 520 507 583 800 800 800 599 800 442
60 500 559 547 588 513 508 548 800 800 800 588 791 450
60 1000 561 554 581 513 508 534 800 800 691 581 615 440
60 2000 564 560 574 511 508 522 800 800 600 574 600 427
4000 60 512 510 524 500 500 500 800 524 800 800 800 456
4000 100 500 500 500 500 500 500 800 500 800 800 800 500
8000 60 505 505 508 500 500 500 800 508 800 800 800 437
8000 100 500 500 500 500 500 500 800 500 800 800 800 500
60 4000 563 561 570 507 506 512 800 800 600 570 600 404
100 4000 600 600 600 600 600 600 800 800 644 600 617 500
60 8000 563 562 568 506 505 507 800 800 600 568 600 383
100 8000 600 600 600 600 600 600 800 800 608 600 602 500
100 10 800 800 800 800 800 800 800 800 800 800 800 754
100 20 800 800 800 800 800 800 800 800 800 800 800 685
10 50 734 687 793 484 437 682 800 800 800 793 800 341
10 100 800 800 800 800 800 800 800 800 800 800 800 800
20 100 800 800 800 799 784 800 800 800 800 800 800 454

of the estimated common factors and common components (i.e., ˆ i F
t ). But using
Theorem 1, it may be possible to obtain these limiting distributions. For example,
the rate of convergence of F
t derived in this paper could be used to examine the
statistical property of the forecast ŷT +1T in Stock and Watson’s √
framework. It
would be useful to show that ŷT +1T is not only a consistent but a T consistent
estimator of yT +1 , conditional on the information up to time T (provided that N
is of no smaller order of magnitude than T ). Additional asymptotic results are
currently being investigated by the authors.
The foregoing analysis has assumed a static relationship between the observed
data and the factors. Our model allows Ft to be a dependent process, e.g,
A
LFt = t , where A
L is a polynomial matrix of the lag operator. However,
we do not consider the case in which the dynamics enter into Xt directly. If the
212 jushan bai and serena ng

method developed in this paper is applied to such a dynamic model, the esti-
mated number of factors gives an upper bound of the true number of factors.
Consider the data generating process Xit = ai Ft + bi Ft−1 + eit . From the dynamic
point of view, there is only one factor. The static approach treats the model as
having two factors, unless the factor loading matrix has a rank of one.
The literature on dynamic factor models is growing. Assuming N is fixed,
Sargent and Sims (1977) and Geweke (1977) extended the static strict factor
model to allow for dynamics. Stock and Watson (1998) suggested how dynamics
can be introduced into factor models when both N and T are large, although
their empirical applications assumed a static factor structure. Forni et al. (2000a)
further allowed Xit to also depend on the leads of the factors and proposed a
graphic approach for estimating the number of factors. However, determining
the number of factors in a dynamic setting is a complex issue. We hope that
the ideas and methodology introduced in this paper will shed light on a formal
treatment of this problem.

Dept. of Economics, Boston College, Chestnut Hill, MA 02467, U.S.A.;

[email protected]
and
Dept. of Economics, Johns Hopkins University, Baltimore, MD 21218, U.S.A.;
[email protected]
Manuscript received April, 2000; final revision received December, 2000.

APPENDIX

To prove the main results we need the following lemma.

Lemma 1: Under Assumptions A–C, we have for some M1 < , and for all N and T ,

T 
T
(i) T −1 )N
s t2 ≤ M1 ,
s=1 t=1
    2 

T T 
 
N 
(ii) E T −1 N −1/2 et 0 2 = E T −1 N
−1/2
eit 0i 
 ≤ M1 
t=1 t=1 i=1
  2 
−2

T 
T
−1

N
(iii) E T N Xit Xis ≤ M1 ,
t=1 s=1 i=1
 
 
N T 
(iv) E

N T  −1/2
e 0
it i  ≤ M1 .

i=1 t=1

Proof: Consider (i). Let 8
s t = )N
s t/)N
s s)N
t t1/2 . Then 8
s t ≤ 1. From
)N
s s ≤ M,


T 
T 
T 
T
T −1 )N
s t2 = T −1 )N
s s)N
t t8
s t2
s=1 t=1 s=1 t=1


T 
T
≤ MT −1 )N
s s)N
t t1/2 8
s t
s=1 t=1


T 
T
= MT −1 )N
s t ≤ M 2
s=1 t=1
 approximate factor models 213

by Assumption C2. Consider (ii).

 2
 −1/2 N  1 N  N
 1 N  N
E
N e 0
it i  =
E
eit ejt  0i 0j ≤ ¯ 2 *  ≤ ¯ 2 M
i=1
N i=1 j=1 N i=1 j=1 ij

by Assumptions B and C3. For (iii), it is sufficient to prove EXit 4 ≤ M1 for all
i t. Now EXit 4 ≤
8E
0i Ft0 4 + 8Eeit 4 ≤ 8 ¯ 4 EFt0 4 + 8Eeit 4 ≤ M1 for some M1 by Assumptions A, B, and C1. Finally


for (iv),
 2
 N  T  1  N N  T  T

E
N T  −1/2
e 0
it i  =
E
eit ejs  0i 0j
i=1 t=1
N T i=1 j=1 t=1 s=1

1  N  N  T  T
≤ ¯ 2 *  ≤ ¯ 2 M
N T i=1 j=1 t=1 s=1 ij ts

by Assumption C4.

Proof of Theorem 1: We use the mathematical identity F
k = N −1 X  k and 

k = T −1 X  F k .

From the normalization F k F k /T = Ik , we also have T −1 Tt=1 F tk 2 = Op
1. For H k =
 

k 0 0 0

F F /T 
  /N , we have

 
T 
T 
T 
T
F
tk − H k Ft0 = T −1 F sk )N
s t + T −1 F sk >st + T −1 F sk ?st + T −1 F sk @st where
s=1 s=1 s=1 s=1

es et
>st = − )N
s t
N
 
?st = Fs0 0 et /N 
 
@st = Ft0 0 es /N = ?ts 

Note that H k depends on N and T . Throughout, we will suppress this dependence to simplify the
notation. We also note that H k  = Op
1 because H k  ≤ F k F k /T 1/2 F 0 F 0 /T 1/2 0 /N  and
  

each of the matrix norms is stochastically bounded by Assumptions A and B. Because
x + y + z +

u2 ≤ 4
x2 + y 2 + z2 + u2  F
tk − H k Ft0 2 ≤ 4
at + bt + ct + dt , where


 T 2
 k 
at = T −2  F )N
s t 
 s 
s=1
 T 
  k 2
bt = T −2  >st  
F
 s 
s=1
 T 
  k 2
ct = T −2  ?st  
F
 s 
s=1
 T 
  k 2
dt = T −2  @st  
F
 s 
s=1

T 
It follows that
1/T  t=1 F
tk − H k Ft0 2 ≤ 1/T Tt=1
at + bt + ct + dt .

T k T 
Now  s=1 Fs )N
s t ≤
s=1 F s   ·
s=1 )N2
s t. Thus,
2 k 2 T

 T    T T 

T  2 
T −1 at ≤ T −1 T −1 F sk  · T −1 )N
s t2
t=1 s=1 t=1 s=1

= Op
T −1 

by Lemma 1(i).
214 jushan bai and serena ng

For bt , we have that

 T
T 
2

T  
bt = T −2  F k >st 
 s 
t=1 t=1 s=1


T 
T 
T

=T −2
F sk F uk >st >ut
t=1 s=1 u=1
 
T  2 1/2

T 
T
 k k 2 1/2 −2 
T  T
≤ T −2 F s F u T >st >ut
s=1 u=1 s=1 u=1 t=1
 
T  2 1/2
  2 T 
T  T
≤ T −1 F sk  · T −2 >st >ut 
s=1 s=1 u=1 t=1
T 2
T T 2 4
From E
t=1 >st >ut  = E
t=1 v=1 >st >ut >sv >uv  ≤ T max s t E>st  and
 
1  N 4
E>st 4 = 2 E N −1/2
eit eis − E
eit eis  ≤ N −2 M
N i=1

by Assumption C5, we have

 
T
T2 T
bt ≤ Op
1 · = O p 
t=1
N2 N

T −1 Tt=1 bt = Op
N −1 . For ct , we have
 T   T 2
  k 2   k 0 0 
ct = T −2   −2  
 Fs ?st  = T  Fs Fs  et /N 
s=1 s=1
  

T 
T
≤ N −2 et 0 2 T −1
F sk 2 T −1 Fs0 2
s=1 s=1

= N −2 et 0 2 Op
1

It follows that

T   0 2


T 
T −1 ct = Op
1N −1 T −1  e√t   = Op
N −1 
 N
t=1 t=1

by Lemma 1(ii). The term dt = Op
N −1  can be proved similarly. Combining these results, we have

T −1 Tt=1
at + bt + ct + dt  = Op
N −1  + Op
T −1 .
To prove Theorem 2, we need additional results.

Lemma 2: For any k, 1 ≤ k ≤ r, and H k defined as the matrix in Theorem 1,

V
k F
k  − V
k F 0 H k  = Op
CN−1T 

Proof: For the true factor matrix with r factors and H k defined in Theorem 1, let MF0H =
I − PF0H denote the idempotent matrix spanned by null space of F 0 H k . Correspondingly, let MF
k =
IT − F
k
F
k F
k −1 F
k = I − PF
k . Then
 


N
V
k F
k  = N −1 T −1 X i MF
k X i 
i=1


N
V
k F 0 H k  = N −1 T −1 X i MF0H X i 
t=1


N
V
k F
k  − V
k F 0 H k  = N −1 T −1 X i
PF0H − PF
k X i 
i=1
 approximate factor models 215

Let Dk = F
k F
k /T and D0 = H k F 0 F 0 H k /T . Then

  


k
k −1  k 0 0 k −1

F F  H F F H  
PF
k − PF0H = T −1 F
k F
k − T −1 F 0 H k Hk F 0
T T
  
= T −1 F
k Dk−1 F
k − F 0 H k D0−1 H k F 0 
 
= T −1 
F
k − F 0 H k + F 0 H k Dk−1
F
k − F 0 H k + F 0 H k  − F 0 H k D0−1 H k F 0 
 
= T −1 
F
k − F 0 H k Dk−1
F
k − F 0 H k  +
F
k − F 0 H k Dk−1 H k F 0
 
+ F 0 H k Dk−1
F
k − F 0 H k  + F 0 H k
Dk−1 − D0−1 H k F 0 

N
Thus, N −1 T −1 i=1 X i
PF
k − PF0H X i = I + II + III + IV . We consider each term in turn.


N 
T 
T
 
I = N −1 T −2
F
tk − H k Ft0  Dk−1
F
sk − H k Fs0 Xit Xis
i=1 t=1 s=1

 1/2
T  2 1/2

T 
T
  
T  
N
≤ T −2 
F
tk − H k Ft0  Dk−1
F
sk − H k Fs0 2 · T −2 N −1 Xit Xis
t=1 s=1 t=1 s=1 i=1

 

T

≤ T −1 Ftk − H k Ft0 2 · Dk−1  · Op
1 = Op
CN−2T 
t=1

by Theorem 1 and Lemma 1(iii). We used the fact that Dk−1  = Op
1, which is proved below.


N 
T 
T
 
II = N −1 T −2
F
tk −H k Ft0  Dk−1 H k Fs0 Xit Xis
i=1 t=1 s=1

 1/2
T  2 1/2

T 
T
  
T  
N
≤ T −2 F
tk −H k Ft0 2 ·H k Fs0 2 ·Dk−1 2 · T −2 N −1 Xit Xis
t=1 s=1 t=1 s=1 i=1

 1/2  1/2
T
  T

≤ T −1 F
tk −H k Ft0 2 ·Dk−1 · T −1 H k Fs0 2 ·Op
1
t=1 s=1

 1/2

T

= T −1 F
tk −H k Ft0 2 ·Op
1 = Op
CN−1T 
t=1

It can be verified that III is also Op
CN−1T .


N 
T 
T
 
IV = N −1 T −2 Ft0 H k
Dk−1 − D0−1 H k Fs0 Xit Xis
i=1 t=1 s=1
 2

N 
T

≤ Dk−1 − D0−1 N −1 T −1 H k Ft0  · Xit 
i=1 t=1

= Dk−1 − D0−1  · Op
1


where Op
1 is obtained because the term is bounded by H k 2
1/T  Tt=1 Ft0 2
1/N T ×
 N T 2 2
i=1 t=1 Xit  , which is Op
1 by Assumption A and EXit  ≤ M. Next, we prove that Dk − D0  =
216 jushan bai and serena ng

Op
CN−1T . From

F
k F
k H k F 0 F 0 H k
  

Dk − D 0 = −
T T
T
k
k  
=T −1

Ft Ft − H k Ft0 Ft0 H k 
t=1


T
 
=T −1

F
tk − H k Ft0 
F
tk − H k Ft0 
t=1


T
  
T
 
+ T −1
F
tk − H k Ft0 Ft0 H k + T −1 H k Ft0
F
tk − H k Ft0  
t=1 t=1
 1/2  1/2

T
 
T
 
T

Dk − D0  ≤ T −1 F
tk − H k Ft0 2 + 2 T −1 F
tk − H k Ft0 2 · T −1 H k Ft0 2
t=1 t=1 t=1

= Op
CN−2T  + Op
CN−1T  = Op
CN−1T 


Because F 0 F 0 /T converges to a positive definite matrix, and because rank
H k  = k ≤ r, D0
k × k
converges to a positive definite matrix. From Dk − D0  = Op
CN−1T , Dk also converges to a positive
definite matrix. This implies that Dk−1  = Op
1. Moreover, from Dk−1 − D0−1 = Dk−1
D0 − Dk D0−1 we
have Dk−1 − D0−1  = Dk − D0 Op
1 = Op
CN−1T . Thus IV = Op
CN−1T .

Lemma 3: For the matrix H k defined in Theorem 1, and for each k with k < r, there exists a *k > 0
such that

plim inf V
k F 0 H k  − V
r F 0  = *k 

N  T →

Proof:

N
V
k F 0 H k  − V
r F 0  = N −1 T −1 X i
PF0 − PF0H X i
i


N
= N −1 T −1
F 0 0i + ei 
PF0 − PF0H 
F 0 0i + ei 
i−1


N
 
= N −1 T −1 0i F 0
PF0 − PF0H F 0 0i
i=1


N
+ 2N −1 T −1 ei
PF0 − PF0H F 0 0i
i=1


N
+ N −1 T −1 ei
PF0 − PF0H ei
i=1

= I + II + III
0 0
First, note that P − P ≥ 0. Hence, III ≥ 0. For the first two terms,
F FH


 
N

I = tr T −1 F 0
PF0 − PF0H F 0 N −1 0i 0i
i=1

0 0
 k
 −1 k 0 0  
F F F F H0 0 k
H F 0 F 0H k H F F N

= tr − · N −1 0i 0i
T T T T i=1

  −1   
→ tr F − F H0k H0k k
F H0 H0k F · D

= tr
A·D
 approximate factor models 217
 
where A = F − F H0k
H0k F H0k −1 H0k F and H0k is the limit of H k with rank
H0k  = k < r. Now
A = 0 because rank
F  = r (Assumption A). Also, A is positive semi-definite and D > 0 (Assumption
B). This implies that tr
A · D > 0.

Remark: Stock and Watson (1998) studied the limit of H k . The convergence of H k to H0k holds
jointly in T and N and does not require any restriction between T and N .
Now


N 
N
II = 2N −1 T −1 ei PF0 F 0 0i − 2N −1 T −1 ei PF0H F 0 0i 
i=1 i=1

Consider the first term.

   
 −1 −1 
N   −1 −1 
N 
T 
0 0 
N T e  0 0 0
P F = N T e F
 i F i  it t i
i=1 i=1 t=1
 1/2  T 
 2 1/2
T
1  1 N 
≤ T −1 Ft0 2 ·√ T −1  √
 N e 0
it i 
t=1 N t=1 i=1

1
= Op √ 
N

√
√ equality follows from Lemma 1(ii). The second term is also Op
1/ N , and hence II =
The last
Op
1/ N  → 0.

Lemma 4: For any fixed k with k ≥ r, V
k F
k  − V
r F
r  = Op
CN−2T ).

Proof:

V
k F
k  − V
r F
r  ≤ V
k F
k  − V
r F 0  + V
r F 0  − V
r F
r 

≤ 2 max V
k F
k  − V
r F 0 

r≤k≤kmax

Thus, it is sufficient to prove for each k with k ≥ r,

(10) V
k F
k  − V
r F 0  = Op
CN−2T 

Let H k be as defined in Theorem 1, now with rank r because k ≥ r. Let H k+ be the generalized
inverse of H k such that H k H k+ = Ir . From X i = F 0 0i + ei , we have X i = F 0 H k H k+ 0i + ei . This
implies

X i = F
k H k+ 0i + ei −
F
k − F 0 H k H k+ 0i

= F
k H k+ 0i + ui 

where ui = ei −
F
k − F 0 H k H k+ 0i .
218 jushan bai and serena ng

Note that

N
V
k F
k  = N −1 T −1 ui MF
k ui 
i=1


N
V
r F 0  = N −1 T −1 ei MF0 ei 
i=1


N
V
k F
k  = N −1 T −1
ei −
F
k − F 0 H k H k+ 0i  MF
k
ei −
F
k − F 0 H k H k+ 0i 
i=1


N 
N

= N −1 T −1 ei MF
k ei − 2N −1 T −1 0i H k+
F
k − F 0 H k  MF
k ei
i=1 i=1


N
0
+ N −1 T −1
H k+
F
k − F 0 H k  MF
k
F
k − F 0 H k H k+ 0i
i
i=1

= a + b + c

Because I − M is positive semi-definite, x MF
k x ≤ x x. Thus,

k
F


N

c ≤ N −1 T −1 0i H k+
F
k − F 0 H k 
F
k − F 0 H k H k+ 0i
i=1
 

T
 
N
≤T −1
F
tk − H k Ft0 2 · N −1  0i 2 H k+ 2
t=1 i=1

= Op
CN−2T  · Op
1

by Theorem 1. For term b, we use the fact that tr
A ≤ rA for any r × r matrix A. Thus
  

N
b = 2T −1 tr H k+
F
k − F 0 H k  MF
k N −1 ei 0i
i=1
 k   
 F
− F 0 H k   1  N 
≤ 2rH k+  · 
 √ ·√
  e i 0
i 
T T N i=1
   T 
 2 1/2
T 1/2
1 1  N 

≤ 2rH k+  · T −1 F
tk − H k Ft0 2 ·√  √1 e 0

N T  N i=1
it i 
t=1 t=1

1
= Op
CN−1T  · √ = Op
CN−2T 
N
by Theorem 1 and Lemma 1(ii). Therefore,

N
V
k F
k  = N −1 T −1 ei MF
k ei + Op
CN−2T 
i=1

Using the fact that V
k F
k  − V
r F 0  ≤ 0 for k ≥ r,

1  N
1  N
(11) 0 ≥ V
k F
k  − V
r F 0  = ei PF
k ei − e P 0 e + Op
CN−2T 
N T i=1 N T i=1 i F i
Note that
1  N
 N

ei PF0 ei ≤ 
F 0 F 0 /T −1  · N −1 T −2 ei F 0 F 0 ei
N T i=1 i=1

N 
 2
 
N 
= Op
1T −1 N −1  T
−1/2
Ft0 eit  −1 −2
 = Op
T  ≤ Op
CN T 
i=1 t=1
 approximate factor models 219

by Assumption D. Thus


N
0 ≥ N −1 T −1 ei PF
k ei + Op
CN−2T 
i=1

N
This implies that 0 ≤ N −1 T −1 i=1 ei PF
k ei = Op
CN−2T . In summary

V
k F
k  − V
r F 0  = Op
CN−2T 

Proof of Theorem 2: We shall prove that limN  T → P
PC
k < PC
r = 0 for all k = r and
k ≤ kmax. Since

PC
k − PC
r = V
k F
k  − V
r F
r  −
r − kg
N  T 

it is sufficient to prove P V
k F
k  − V
r F
r  <
r − kg
N  T  → 0 as N  T → . Consider k < r.
We have the identity:

V
k F
k  − V
r F
r  = V
k F
k  − V
k F 0 H k  + V
k F 0 H k  − V
r F 0 H r 

+ V
r F 0 H r  − V
r F
r 

Lemma 2 implies that the first and the third terms are both Op
CN−1T . Next, consider the second
term. Because F 0 H r and F 0 span the same column space, V
r F 0 H r  = V
r F 0 . Thus the second
term can be rewritten as V
k F 0 H k  − V
r F 0 , which has a positive limit by Lemma 3. Hence,
P PC
k < PC
r → 0 if g
N  T  → 0 as N  T → . Next, for k ≥ r,

P PC
k − PC
r < 0 = P V
r F
r  − V
k F
k  >
k − rg
N  T 

By Lemma 4, V
r F
r  − V
k F
k  = Op
CN−2T . For k > r
k − rg
N  T  ≥ g
N  T , which converges
to zero at a slower rate than CN−2T . Thus for k > r P PC
k < PC
r → 0 as N  T → .

Proof of Corollary 1: Denote V
k F
k  by V
k for all k. Then

IC
k − IC
r = lnV
k/V
r +
k − rg
N  T 

For k < r, Lemmas 2 and 3 imply that V
k/V
r > 1 + 0 for some 0 > 0 with large probability
for all large N and T . Thus lnV
k/V
r ≥ 0 /2 for large N and T . Because g
N  T  → 0, we
have IC
k − IC
r ≥ 0 /2 −
r − kg
N  T  ≥ 0 /3 for large N and T with large probability. Thus,
P IC
k − IC
r < 0 → 0. Next, consider k > r. Lemma 4 implies that V
k/V
r = 1 + Op
CN−2T .
Thus lnV
k/V
r = Op
CN−2T . Because
k − rg
N  T  ≥ g
N  T , which converges to zero at a
slower rate than CN−2T , it follows that
   
P IC
k − IC
r < 0 ≤ P Op CN−2T + g
N  T  < 0 → 0

Proof of Corollary 2: Theorem 2 is based on Lemmas 2, 3, and 4. Lemmas 2 and 3 are still
valid with F k replaced by G
k and CN T replaced by C
N T . This is because their proof only uses the
convergence rate of F
t given in (5), which is replaced by (8). But the proof of Lemma 4 does make
use of the principle component property of F
k such that V
k F
k  − V
r F 0  ≤ 0 for k ≥ r, which is
not necessarily true for G
k . We shall prove that Lemma 4 still holds when F
k is replaced by G
k and
CN T is replaced by C N T . That is, for k ≥ r,
 −2 
(12)
k  − V
r G
V
k G 

r  = Op C
NT

Using arguments similar to those leading to (10), it is sufficient to show that

 −2 
(13)
k  − V
r F 0  = Op C
V
k G 
NT
220 jushan bai and serena ng

Note that for k ≥ r,

(14)
k  ≤ V
r G
V
k F
k  ≤ V
k G
r 

The first inequality follows from the definition that the principal component estimator gives the
smallest sum of squared residuals, and the second inequality follows from the least squares property
that adding more regressors does not increase the sum of squared residuals. Because C 2 ≤ C 2 , we
NT NT
can rewrite (10) as
 −2 
(15) 
V
k F
k  − V
r F 0  = Op C NT

It follows that if we can prove

 −2 
(16)
r  − V
r F 0  = Op C
V
r G 
NT

then (14), (15), and (16) imply (13). To prove (16), we follow the same arguments as in the proof of
Lemma 4 to obtain

r  − V
r F 0  = 1  N
1  N
 −2 

V
r G ei PG
r ei − e  P 0 e + Op C
N T i=1 N T i=1 i F i NT

r
G
where PG
r = G
r  G

r −1 G

r  ; see (11). Because the second term on the right-hand side is shown in
−1
Lemma 4 to be Op
T , it suffices to prove the first term is Op
C −2 . Now,
NT

N  
1  N

r /T −1  1

r  G


r /T 2 
e G
e P r e ≤ 
G
N T i=1 i G
i N i=1 i

Because H r is of full rank, we have 
Ĝr  Ĝr /T −1  = Op
1 (follows from the same arguments in
proving Dk−1  = Op
1). Next,

N    2    T
1   1  N  T   r 2 N  T
1  
e Ĝr /T 2 ≤  √1 F 0
e  H  + 1 eit2 G r  F 0 2

r − H
N i=1 i
N T i=1  T t=1
t it 
N T i=1 t=1 T t=1 t t

−2  = Op
C
= Op
T −1 Op
1+Op
1Op
C −2 
NT NT

by Assumption D and (8). This completes the proof of (16) and hence Corollary 2.

REFERENCES

Anderson, T. W. (1984): An Introduction to Multivariate Statistical Analysis. New York: Wiley.

Backus, D., S. Forsei, A. Mozumdar, and L. Wu (1997): “Predictable Changes in Yields and
Forward Rates,” Mimeo, Stern School of Business.
Campbell, J., A. W. Lo, and A. C. Mackinlay (1997): The Econometrics of Financial Markets.
Princeton, New Jersey: Princeton University Press.
Chamberlain, G., and M. Rothschild (1983): “Arbitrage, Factor Structure and Mean-Variance
Analysis in Large Asset Markets,” Econometrica, 51, 1305–1324.
Cochrane, J. (1999): “New Facts in Finance, and Portfolio Advice for a Multifactor World,” NBER
Working Paper 7170.
Connor, G., and R. Korajzcyk (1986): “Performance Measurement with the Arbitrage Pricing
Theory: A New Framework for Analysis,” Journal of Financial Economics, 15, 373–394.
(1988): “Risk and Return in an Equilibrium APT: Application to a New Test Methodology,”
Journal of Financial Economics, 21, 255–289.
(1993): “A Test for the Number of Factors in an Approximate Factor Model,” Journal of
Finance, 48, 1263–1291.
 approximate factor models 221

Cragg, J., and S. Donald (1997): “Inferring the Rank of a Matrix,” Journal of Econometrics, 76,
223–250.
Dhrymes, P. J., I. Friend, and N. B. Glutekin (1984): “A Critical Reexamination of the Empir-
ical Evidence on the Arbitrage Pricing Theory,” Journal of Finance, 39, 323–346.
Donald, S. (1997): “Inference Concerning the Number of Factors in a Multivariate Nonparameteric
Relationship,” Econometrica, 65, 103–132.
Forni, M., M. Hallin, M. Lippi, and L. Reichlin (2000a): “The Generalized Dynamic Factor
Model: Identification and Estimation,” Review of Economics and Statistics, 82, 540–554.
(2000b): “Reference Cycles: The NBER Methodology Revisited,” CEPR Discussion Paper
2400.
Forni, M., and M. Lippi (1997): Aggregation and the Microfoundations of Dynamic Macroeconomics.
Oxford, U.K.: Oxford University Press.
(2000): “The Generalized Dynamic Factor Model: Representation Theory,” Mimeo, Univer-
sitá di Modena.
Forni, M., and L. Reichlin (1998): “Let’s Get Real: a Factor-Analytic Approach to Disaggregated
Business Cycle Dynamics,” Review of Economic Studies, 65, 453–473.
Geweke, J. (1977): “The Dynamic Factor Analysis of Economic Time Series,” in Latent Variables in
Socio Economic Models, ed. by D. J. Aigner and A. S. Goldberger. Amsterdam: North Holland.
Geweke, J., and R. Meese (1981): “Estimating Regression Models of Finite but Unknown Order,”
International Economic Review, 23, 55–70.
Ghysels, E., and S. Ng (1998): “A Semi-parametric Factor Model for Interest Rates and Spreads,”
Review of Economics and Statistics, 80, 489–502.
Gregory, A., and A. Head (1999): “Common and Country-Specific Fluctuations in Productivity,
Investment, and the Current Account,” Journal of Monetary Economics, 44, 423–452.
Gregory, A., A. Head, and J. Raynauld (1997): “Measuring World Business Cycles,” Interna-
tional Economic Review, 38, 677–701.
Lehmann, B. N., and D. Modest (1988): “The Empirical Foundations of the Arbitrage Pricing
Theory,” Journal of Financial Economics, 21, 213–254.
Lewbel, A. (1991): “The Rank of Demand Systems: Theory and Nonparametric Estimation,” Econo-
metrica, 59, 711–730.
Mallows, C. L. (1973): “Some Comments on Cp ,” Technometrics, 15, 661–675.
Ross, S. (1976): “The Arbitrage Theory of Capital Asset Pricing,” Journal of Finance, 13, 341–360.
Rubin, D. B., and D. T. Thayer (1982): “EM Algorithms for ML Factor Analysis,” Psychometrika,
57, 69–76.
Sargent, T., and C. Sims (1977): “Business Cycle Modelling without Pretending to Have too
much a Priori Economic Theory,” in New Methods in Business Cycle Research, ed. by C. Sims.
Minneapolis: Federal Reserve Bank of Minneapolis.
Schwert, G. W. (1989): “Tests for Unit Roots: A Monte Carlo Investigation,” Journal of Business
and Economic Statistics, 7, 147–160.
Stock, J. H., and M. Watson (1989): “New Indexes of Coincident and Leading Economic Indica-
tions,” in NBER Macroeconomics Annual 1989, ed. by O. J. Blanchard and S. Fischer. Cambridge:
M.I.T. Press.
(1998): “Diffusion Indexes,” NBER Working Paper 6702.
(1999): “Forecasting Inflation,” Journal of Monetary Economics, 44, 293–335.