Empirical Asset Pricing

Example of GMM

Consider the following regression model:

yt = βxt + t , (1)

where t ∼ iid (0, σ2 ) and E [xt t ] = 0. From the moment condition E [xt t ] = 0, β can be
expressed by
h i−1
β = E x2t E [xt yt ] . (2)

Proof First we replace t with yt − βxt , so

E [xt t ] = E [xt (yt − βxt )]

h i
= E xt yt − βx2t
h i
= E [xt yt ] − βE x2t

= 0.

−1
Therefore, β = E [x2t ] E [xt yt ].

Equation (2) can be consistently estimated by

T
!−1 T
!
1X 1X
β̂ = x2t xt y t . (3)
T t=1 T t=1

This estimator is exactly same with OLS estimator.

[1]
 Let’s assume that xt is not independent of  any longer, E [xt t ] 6= 0. Instead, we find
two instrumental variables z1t and z2t which are independent of t such that

E [z1t t ] = 0. (4)

E [z2t t ] = 0. (5)

From (4), following the previous drivation,

β1 = E [z1t xt ]−1 E [z1t yt ] , (6)

where β1 is the definition of β from the first moment condition (4). Analogously,

β2 = E [z2t xt ]−1 E [z2t yt ] , (7)

where β2 is the definition of β from the first moment condition (4). Further β1 and β2 can
be consistently estimated by

T
!−1 T
!
1X 1X
β̂1 = z1t xt z1t yt . (8)
T t=1 T t=1
T
!−1 T
!
1X 1X
β̂2 = z2t xt z2t yt . (9)
T t=1 T t=1

However, we face the identifcation problem of β definition. We have a unique parameter

in the model (1) but we have two definitions of β in (6) and (7). Since two definitions are
theoretically same, their estimator β̂1 and β̂2 should be same. But we cannot guarantee that
β̂1 and β̂2 are identical each other, since they depend on the sample observations of random
variable z1t , z2t , xt and yt .
Now, our interest is which estimate we will use as an optimal estimator of β given
two consitent estimators. There may be many ideas of constructing the optimal one. For

[2]
example, I suggest the weighted average of two estimators such that

β̂opt = ω β̂1 + (1 − ω) β̂2 ,

where 0 ≤ ω ≤ 1. The reason is that the weighted average of two estimators can reduce the
standard error compare with a single estimator. Let’s extend this idea to a system approach.
First, let    
1 PT 1 PT
T t=1 z1t t  T t=1 z1t (yt − βxt )
ḡ (β) =  P
 
 = P
 ,
 (10)
1 T 1 T
T t=1 z2t t T t=1 z2t (yt − βxt )

where ḡ (β) denotes sample moment conditions. Then we will find β that minimizes a loss
function
J (β) = ḡ (β)0 W ḡ (β) , (11)

where W is an optimal weighting matrix which imposes weight on sample moment conditions
in a systematic way. By optimization theory, the minimization condition is

   0
∂ḡ β̂

∂J (β)   
 = 2  W ḡ β̂ = 0. (12)
∂β β=β̂ ∂β

Hence,

 0   
1 PT 1 PT
t=1 z1t xt  t=1 z1t yt − β̂xt 

∂J (β)  T T
 = −2  P
  W 
∂β    
1 T 1 T
yt − β̂xt
P
t=1 z2t xt t=1 z2t
β=β̂
T T
 0   
1 PT 1 PT 1 PT
T t=1 z1t xt  T t=1 z1t yt − β̂ T t=1 z1t xt 
= −2  P
 
 W P




1 T 1 T 1 T
− β̂
P
T t=1 z2t xt T t=1 z2t yt T t=1 z2t xt
 0    
1 PT 1 PT 1 PT
T t=1 z1t xt   T t=1 z1t yt  T t=1 z1t xt 
= −2  P
 
 W  P
  − β̂
 
 

1 T 1 T 1 PT
T t=1 z2t xt T t=1 z2t yt T t=1 z2t xt
= 0.

[3]
Then,

 0    0  
1 PT 1 PT 1 PT 1 PT
 T t=1 z1t xt  T t=1 z1t xt  T t=1 z1t xt  T t=1 z1t yt 
β̂  P
 
 W P 
 = P
 
 W P
 .

1 T 1 T 1 T 1 T
T t=1 z2t xt T t=1 z2t xt T t=1 z2t xt T t=1 z2t yt

Therefore,

 0  −1  0  
1 PT 1 PT 1 PT 1 PT
 T t=1 z1t xt  T t=1 z1t xt   T t=1 z1t xt  T t=1 z1t yt 
β̂ =  P
 
 W P  




 W P
  .
 (13)
1 T 1 T 1 PT 1 T
T t=1 z2t xt T t=1 z2t xt T t=1 z2t xt T t=1 z2t yt

If W is a true optimal weighting matrix, β̂ is a consistent and efficient estimator of β.This

system approach is a special case of GMM.

[4]