Microstructure of Financial Markets

Assignment M2 MASEF
Marouane IZMAR

Price dynamics in a Glosten-Milgrom model

Consider the multi-period Glosten-Milgrom model, where the security’s true value v can
be high, v H = 1, or low, v L = 0 with initial probability θ0 = 1/2 each. Market makers
are competitive and risk neutral, and do not know v. A single trade comes to the market:
with probability 1 − π, he is a noise trader, who buys or sells one unit with probability
1/2 each; with probability π, he is an informed trader, who knows the realization of v.
We denote θt , the posterior probability, after t trades, that the security’s value is v = 1.

1. In this setup, remind what is the equilibrium strategy of an informed

trader who arrives in the market.
In equilibrium, an informed trader in the multi-period Glosten–Milgrom model will
always buy if the true value of the security is 1 (high) and always sell if the true value is
0 (low). This behavior exploits the trader’s private information and maximizes expected
profit given that market makers set prices based on the conditional expectation of the
asset’s value after observing the order flow.

2. Assuming that the t + 1-th trader sends a buy order, compute the
posterior probability conditional on buy, θt+1 (b) as a function of θt . Explain
why θt+1 (b) is also the price at which the trade occurs.
By total probability,
P (buy) = θt P (buy | v = 1) + (1 − θt )P (buy | v = 0).
Substituting the values:
1 1+π
P (buy | v = 1) = π × 1 + (1 − π) × = ,
2 2
1 1−π
P (buy | v = 0) = π × 0 + (1 − π) × = .
2 2
Thus,
1+π 1−π
P (buy) = θt + (1 − θt ) .
2 2
By Bayes’ rule,
θt P(buy | v = 1) θt 1+π
2
θt+1 (b) = P(v = 1 | buy) = = 1+π .
P(buy) θt 2 + (1 − θt ) 1−π
2

1
 1
Factoring out 2
in numerator and denominator simplifies to

θt (1 + π)
θt+1 (b) = .
θt (1 + π) + (1 − θt )(1 − π)

In the Glosten–Milgrom model, risk-neutral, competitive market makers set the price
of the transaction equal to the expected value of the asset given the order they observe.
Since the asset’s payoff is 1 with probability θt+1 (b) and 0 otherwise, the fair price for a
buy order is exactly

Price = E[v | buy] = θt+1 (b).

Thus, the updated posterior θt+1 (b) is not only the new belief about v = 1 but also
the price at which the buy trade occurs.

3. Similarly, assuming that the t + 1-th trader sends a sell order, compute
the posterior probability conditional on buy, θt+1 (s) as a function of θt . Explain
why θt+1 (s) is also the price at which the trade occurs.

• With probability π, the trader is informed:

– If v = 1, an informed trader buys (so the probability of a sell is 0).

– If v = 0, an informed trader sells (so the probability of a sell is 1).

• With probability 1−π, the trader is a noise trader, who buys or sells with probability
1/2 each, regardless of v.

Therefore:
1 1−π
P(sell | v = 1) = π × 0 + (1 − π) = ,
2 2
1 1−π 1+π
P(sell | v = 0) = π × 1 + (1 − π) = π + = .
2 2 2
By total probability:

P(sell) = θt P(sell | v = 1) + (1 − θt )P(sell | v = 0).

Substituting the expressions above gives:

1−π 1+π
P(sell) = θt + (1 − θt ) .
2 2
By Bayes’ rule:

θt P(sell | v = 1) θt 1−π
2
θt+1 (s) = P(v = 1 | sell) = = 1−π .
P(sell) θt 2 + (1 − θt ) 1+π
2

2
 1
Factoring out 2
from the numerator and denominator yields:

θt (1 − π)
θt+1 (s) = .
θt (1 − π) + (1 − θt )(1 + π)
In the Glosten–Milgrom framework, risk-neutral, competitive market makers set the
transaction price to the expected value of the asset conditional on the observed order.
Because the payoff is 1 with probability θt+1 (s) given a sell order, the fair price at which
the trade takes place is

Price (sell) = E[v | sell] = θt+1 (s).

Hence, the posterior θt+1 (s) serves as both the market’s new belief about v = 1 and
the price at which the sell occurs.

4. Define the new variable,

 
θt
zt = log
1 − θt

and determine the relation between zt+1 , zt , and the trade direction dt+1 , where dt+1 = +1
if the t + 1-th trade is a buy, and dt+1 = −1 if the t + 1-th trade is a sell.
We define
 
θt
zt = ln ,
1 − θt
which is the (log) odds form of the belief θt . We want to find how zt+1 relates to zt
and the trade direction dt+1 ∈ {+1, −1}, where

• dt+1 = +1 if the (t + 1)-th trade is a buy,

• dt+1 = −1 if the (t + 1)-th trade is a sell.

Recall that the “odds” form of Bayes’ rule says

θt+1 θt P(trade | v = 1)
= × ,
1 − θt+1 1 − θt P(trade | v = 0)
where “trade” is either a buy or a sell. Taking the natural logarithm of both sides
gives
 
P(trade | v = 1)
zt+1 = zt + ln .
P(trade | v = 0)

• Buy:
1+π
P(buy | v = 1) 2 1+π
= 1−π = .
P(buy | v = 0) 2
1−π

• Sell:

3
 Sell:
1−π
P(sell | v = 1) 2 1−π
= 1+π = .
P(sell | v = 0) 2
1+π
Hence:
• If dt+1 = +1 (buy),  
1+π
zt+1 = zt + ln .
1−π
• If dt+1 = −1 (sell),
   
1−π 1+π
zt+1 = zt + ln = zt − ln .
1+π 1−π
Because    
1−π 1+π
ln = − ln ,
1+π 1−π
we can write:  
1+π
zt+1 = zt + dt+1 ln ,
1−π
where dt+1 = +1 for a buy and −1 for a sell.

5. DeterminePthe equation for the relation between zt and the cumulative

order flow Dt = ti=1 di .
We start from the result in Question 4, which says that if
 
θt
zt = ln ,
1 − θt
then for each trader t + 1,
 
1+π
zt+1 = zt + dt+1 ln ,
1−π
where dt+1 = +1 if the (t + 1)-th trade is a buy, and dt+1 = −1 if it is a sell.
We want to express zt in terms of z0 and the entire trading history up to time t. Let
us write out the first few steps explicitly: Few steps explicitly:
1. At t = 0:  
1+π
z1 = z0 + d1 ln .
1−π
2. At t = 1:
      
1+π 1+π 1+π
z2 = z1 + d2 ln = z0 + d1 ln + d2 ln .
1−π 1−π 1−π

Simplifying:  
1+π
z2 = z0 + (d1 + d2 ) ln .
1−π

4
 3. At t = 2:
   
1+π 1+π
z3 = z2 + d3 ln = z0 + (d1 + d2 + d3 ) ln .
1−π 1−π

Continuing this pattern up to t, we see that

 
1+π
zt = z0 + (d1 + d2 + · · · + dt ) ln .
1−π
Hence,
 
1+π
zt = z0 + Dt ln .
1−π
 
θt
From the definition zt = ln 1−θt
, we have initially at t = 0:
 
θ0
z0 = ln .
1 − θ0
Putting all this together gives the final expression:
   
θ0 1+π
zt = ln + Dt ln .
1 − θ0 1−π
This shows how the log-odds zt is updated additively by the net order flow Dt .

6. From now on, we take the following parametrization, v = 1 and π =

0.2, and we analyze the dynamics of the true stochastic process Dt , that is
conditional of v = 1. Compute its expectation, E[Dt ], and variance, Σt .
We assume v = 1 and π = 0.2. Recall that:
• With probability π = 0.2, the arriving trader is informed (and will buy for v = 1,
i.e. d = +1).

• With probability 1 − π = 0.8, the trader is noise and buys or sells with probability
1/2 each.
Hence, conditional on v = 1, each trade direction ds takes values +1 or −1 according
to:
1
ds = +1 with probability π + (1 − π) = 0.2 + 0.4 = 0.6.
2
1
ds = −1 with probability (1 − π) = 0.4.
2

P(ds = +1) = 0.6, P(ds = −1) = 0.4.

E[ds ] = (+1) × 0.6 + (−1) × 0.4 = 0.2.

5
 Since d2s = 1 in either case,

E[d2s ] = 1, so Var(ds ) = E[d2s ] − (E[ds ])2 = 1 − (0.2)2 = 0.96.

and
t
X
Dt = ds .
s=1

If we assume the ds are i.i.d. given v = 1, then:

• Expectation:
t
X
Dt = E[Dt ] = E[ds ] = t × 0.2 = 0.2t.
s=1

• Variance:
t
X
Σt = Var(Dt ) = Var(ds ) = t × 0.96 = 0.96t.
s=1

Thus, under v = 1 and π = 0.2, the random process Dt has

Dt = 0.2t, Σt = 0.96t.

7. What is the value DT such that DT > D implies θT > 0.99? How many
trades does it take for this threshold to be crossed on average, that is such
that Dt > D?
Recall that in the Glosten–Milgrom setup (assuming prior θ0 = 0.5),
   
θt 1+π
zt = ln = Dt ln ,
1 − θt 1−π
because z0 = 0 when θ0 = 0.5.
We want θt > 0.99. Since
e zt
θt = ,
1 + ezt
the condition θt > 0.99 is equivalent to
ezt θt 0.99
> 0.99 ⇐⇒ = ezt > = 99.
1 + e zt 1 − θt 0.01
Hence,

zt > ln(99).
1+π


Substituting zt = Dt ln 1−π . With π = 0.2, we have
   
1+π 1.2
ln = ln = ln(1.5) ≈ 0.40536.
1−π 0.8

6
 Thus,

ln(99)
Dt ln(1.5) > ln(99) ⇐⇒ Dt > .
ln(1.5)
Numerically, ln(99) ≈ 4.59512. So

ln(99) 4.59512
≈ ≈ 11.34.
ln(1.5) 0.40536
Because Dt must be an integer, the smallest integer D satisfying this is

D = 12.
Therefore, if Dt ≥ 12, then θt > 0.99.
Under v = 1 and π = 0.2, each trader is: - Informed (buys) with probability 0.2.
- Noise (50/50 buy or sell) with probability 0.8.
Hence each trade direction ds has

P(ds = +1) = 0.2 + 0.8 × 0.5 = 0.6, P(ds = −1) = 0.4.

So E[ds ] = 0.6 − 0.4 = 0.2. Then the expected cumulative order flow after t trades is

Dt = E[Dt ] = t × 0.2.
We want to find the first t such that Dt exceeds D = 12. Solve

0.2 t = 12 =⇒ t = 60.
Hence, on average, we expect to cross the threshold D = 12 after about t = 60 trades.

8. Using the normal approximation provided by the central limit theorem

for Dt , derive and plot as a function of t, the probability:

P r[θt > 0.99]

Determine graphically how many trades does it takes for the former probability to
exceed 95%?
We continue under the same assumptions as before:

• ds = +1 with probability 0.6 (informed buy or noise buy) and ds = −1 with

probability 0.4,

• hence Dt = ts=1 ds has mean 0.2t and variance 0.96t.

P

• We start with prior θ0 = 0.5, so z0 = ln 0.5

0.5


= 0.

Recall the log-odds update:

   
θt 1+π
zt = ln = Dt ln with π = 0.2.
1 − θt 1−π

7
 Because
1 + 0.2
= 1.5,
1 − 0.2
we get

zt = Dt ln(1.5).
We want

ezt 0.99
θt > 0.99 ⇐⇒ > 0.99 ⇐⇒ e zt > = 99 ⇐⇒ zt > ln(99).
1 + e zt 0.01
Since zt = Dt ln(1.5), the condition becomes

ln(99) 4.5951
Dt > ≈ ≈ 11.34.
ln(1.5) 0.4054
So the discrete threshold is Dt ≥ 12.
Under i.i.d. ds , by the Central Limit Theorem, for large t,

Dt ≈ N (µt , σt2 ),
where µt = 0.2t, σt2 = 0.96t.
We have
 
11.34 − 0.2t
P (θt > 0.99) = P (Dt > 11.34) ≈ 1 − Φ √ ,
0.96t
where Φ is the standard normal CDF.
Hence, if you plot
 
11.34 − 0.2t
p(t) = 1 − Φ √
0.96t
as a function of t, you obtain the approximate probability that θt exceeds 0.99.
We want

   
11.34 − 0.2t 11.34 − 0.2t
P (Dt > 11.34) > 0.95 ⇐⇒ 1−Φ √ > 0.95 ⇐⇒ Φ √ < 0.05.
0.96t 0.96t

Since Φ−1 (0.05) ≈ −1.645, we set

11.34 − 0.2t
√ = −1.645.
0.96t
Rearranging,
√
0.2t − 11.34 = 1.645 0.96t.
Solving numerically gives

t ≈ 158.

8
 In other words, once t is around 158 trades, the normal approximation predicts
that the probability θt > 0.99 (i.e. Dt > 11.34) exceeds 95%.