FINA 4103

Class Note 13: Measuring Liquidity

Jun Aoyagi
HKUST

Fall 2023

0 / 36
Measure of Liquidity: What Theories Suggest?

I We learned canonical models that explain bid-ask spread

I information asymmetry (Glosten-Milgrom)
I order-processing cost (Stoll and Huang)
I inventory cost (Stoll)
I And the price impact of order flow
I again, information asymmetry matters (Kyle)
I As a trader, you are concerned about
I short-run cost (bid-ask spread)
I transient/long-run price impact

I In reality, what can we observe?

1 / 36
Actual Measure of Liquidity
I In general, information and utility-related variables are not
observable
I Who is informed/noise? How much informed?
I How risk averse dealers are?
I What we can observe is the trading cost
I Reverse engineer and estimate frictions

2 / 36
Measures of Trading Costs

I Some explicit costs are easy to observe

I Ex, dealer commissions, trading fees, etc
I Steadily decreasing since 2000s

I What about the implicit cost of trading?

I Implicit: costs arising from illiquidity
I Also declined ∼ 2007 but shot up afterwards

I Basic idea:

Illiquidity cost = Execution price - Benchmark

I benchmark is the price obtained in perfectly liquid market

3 / 36
Measures of Trading Costs

I Depending on “what if” situation, we have several measures of

spreads
1. Quoted spread
2. Effective spread
3. Realized spread
I Other trading cost measures
1. VWAP
2. Price impact measure
3. Amihud liquidity ratio
4. Roll’s measure

I Measure of information: PIN

4 / 36
Measures Based on Quotes Information

5 / 36
Quoted Spread

I From LOB data, we observe (N)BBO, i.e., A and B.

I The quoted bid-ask spread (“the” bid-ask spread):

S= A−B

I Sometimes use relative spread: normalized by the mid-quote

S S
s= = .
m ( A + B)/2

I Quoted spread measures the implicit cost for a small-size

round-trip trade

6 / 36
Determinant of Quoted Spread

I The GM model implies S is determined by

I µ: information-to-noise (signal-to-noise) ratio
I σ: value uncertainty
I δ, v̄: dealers’ belief

I Observing S, cannot separate these parameters (as well as other

factors)
I We may estimate how much information the order flow contains
I called PIN (probability of informed trading) measure
I will see later

7 / 36
Quoted Spread for a Large Order

I What if a trade has a larger size, |q| > 1?

I the order must walk up (down) the LOB, being executed at worse
prices

I Use average bid and ask prices to execute q

S̄ = Ā(q) − B̄(q)

I Ā(q) is the average ask price the order must pay

I In both cases, the quoted spread reflects the liquidity available at a

given point of time for hypothetical transactions

8 / 36
Data Issues in Quoted Spread

I For the quoted spread, need (N)BBO data at various points in time
I For the average quoted spread, need information about depth and
prices beyond (N)BBO
I this information is not always readily available

I Practitioners often measure the costs based on transaction price

data

9 / 36
Effective Spread

I The quoted spread is for a hypothetical round-trip trade at a

given point in time

I The actual execution price you pay/receive will be different from

BBO spread
I Effective spreads measure the actual trading cost
I A market buy trades at average price of $75.5, while the prevailing
mid-quote is $75.45
I What is the cost relative to the perfectly liquid market?

10 / 36
Effective Spread

Effective spread
Effective spread is the difference between the price at which a market
order is executed and the mid-quote on the market the instant before.

I Let d ∈ {+1, −1} be the direction of order

I If d is executed at price p and mid-quote is m

Se = d ( p − m )

I Relative effective spread: se = Se /m

I m: benchmark when market is perfectly liquid (i.e., no spread)

I For |q| > 1, use the average execution price p̄(q) − m

11 / 36
Effective Spread: Implications

I Se is also called a “slippage” measure

I your order triggers change in the price
I order’s impact on the price
I Based on the past information, capturing retrospective market
liquidity
I unlike the quoted spread, which is liquidity for incoming (future)
orders
I If average over many trades, Se measures market liquidity
I how each order impact the price and causes deviation of post-trade
price from the mid quote

12 / 36
Issues in Using Effective Spread

Split orders
Consider a large market order split into multiple small ones and
traded gradually. We want to compare the cost of split orders with that
of trading all at once.

I If you are a trader, easy to compute the effective spread for split
orders
I Compute average execution price; and
I m just before the first split order

I What if you are a researcher or policy maker?

13 / 36
Issues in Using Effective Spread (cont’d)

Misalignment in data
We see a transaction at price $98.88 at a moment when the best bid and
ask are $98.85 and $98.92.

I The trade might happen a few seconds earlier when b = 98.88 and
a = 98.92 (so its “sell” initiated)
I but opposite is also possible; data may not tell you the trading
direction
I Lee and Ready (1991): misalignment in data
I quote recording is generally faster than the trade recording
I delays up to five seconds are common

I They proposed the algorithm to classify (but not perfect)

14 / 36
Profits for Liquidity Providers (Market Makers)

I Quoted and effective spreads = cost for liquidity takers

I What about the cost (or profit) of liquidity providers?
I Naive prediction: taker’s cost = maker’s profit
I but need to incorporate timing

15 / 36
Profits for Liquidity Providers (Market Makers)

I A dealer buys at $100 from her customer when B = $100 and

A = $102
I To unwind this position, she may sell it at prevailing ask at
A = $102
I leading to $2 profits per share
I A − B would be a good measure
I But what if the prices change to B0 = $99 and A0 = $100?
I e.g., due to information updates

I If the price moves in the direction of liquidity-taking orders,

effective (and quoted) spreads overestimate maker’s profit

16 / 36
Realized Spread

Realized Spread
The realized (half) spread is the difference between (i) the execution
price and (ii) the mid-price at some time after the transaction.

I Let the market incorporate information into the quotes

I Estimate the spread by using mt+∆ as benchmark
I Choice of ∆ depends on
I data frequency and research objective
I market characteristics

17 / 36
Realized Spread

I Formally, it is defined as

Sr = d t ( p t − m t + ∆ )

I dt : trade direction
I pt : execution price
I mt : mid-point quote

I Trade at t and unwind at t + ∆ (in perfectly liquid market)

18 / 36
Realized Spread: Decomposition
I We can re-write it as
Se
z }| {
Sr = d t ( p t − m t ) − d t ( m t + ∆ − m t )

∴ E[Sr ] = E[Se ] − E[dt (mt+∆ − mt )]

I Sr (profit for dealers) is smaller than Se (cost for taker) if

E[dt (mt+∆ − mt )] > 0

I Change in the mid-price is positively correlated with order
direction
I “buy” leads to upward revision in belief (m
t+∆ > mt )

I What if E[Se ] is very small?

19 / 36
Quoted, Effective, and Realized Spreads

I In 2005, SEC rule 605 mandates reporting of monthly statistics by

exchanges
I called “Dash-5” reports

I Include average effective spreads, realized spreads, and spreads

for various order sizes.

20 / 36
Dash-5 Reporting of Microsoft in 2005

21 / 36
Other Measures of Trading Costs

22 / 36
VWAP

I Let’s say we don’t have quote data

I cannot compute and use the mid-point quote as benchmark

I Use the average transaction prices over a day

I Volume-weighted average price (VWAP)

$ volume of trading
VWAP = = ∑ wt p t
# of shares traded t∈ T

with
|qt |
wt =
∑t∈ T | qt |

I qt and pt : size and price of tth trade

23 / 36
VWAP

I Evaluate your trade at your trading price P and VWAP

I involves comparison with effective and realized spreads
I Obvious problem arises when your order has a significant share
I say, your order is the only order in the day

I Also, when brokers are evaluated relative to VWAP, they have

incentive to game with VWAP

24 / 36
Price Impact Measure

I Use data on mid-prices and order imbalances

I The price impact estimation:

mt+∆ − mt = λqt + ut (1)

I ∆: appropriate lag
I qt : order imbalance (buy - sell volume)
I ut : estimation error
I λ̂ measures how order pressures affect the mid-quotes
I c.f., Kyle (1985) model

25 / 36
Price Impact Measure: Estimation Result

I Stoll (2000, JF): estimation in NYSE/AMSE and Nasdaq

I λ > 0 for 95% of stocks (5% significance)
I λ = 0.75% for small-cap stocks
I λ = 0.5% for large-cap stocks
I Hasbrouck (2007): modified the estimation using trading volume
I when trading direction is not observable

26 / 36
Amihud Liquidity Measure

I Amihud (2002) introduced the “illiquidity” ratio

|r t |
It =
Volt
I The absolute return |rt | relative to volume Volt over a given period
I Measuring changes in return when additional volume is traded
I Implication is similar to the price impact regression in (1)
I m
t+∆ − mt is replaced by return rt
I qt is replaced by volume Volt
I Can compute without using quote/order direction data

27 / 36
Some Other Measures

I No-trading: Lesmond, Ogden, and Trzcinka (1999)

I fraction of no-trade days (or time with no price changes) as a
measure of illiquidity
I Implementation shortfall: Perold (1988)
I brokers sometimes delay customer’s order excessively
I derive measure incorporating the delays

28 / 36
Measuring Information Frictions

I Glosten and Milgrom (1985) and Kyle (1985) are based on

information asymmetry
I informed-to-noise ratio matters
I caveat: we cannot directly observe informed/noise

I But order flow is observable

I Can we reverse estimate it from order flow?
I Easley, Hvidkjaer, and O’Hara (2002)

29 / 36
Information Measure: Event Uncertainty

I Each period has two possibilities:

I α: asset fundamentals change (v → v ± σ)
I 1 − α: nothing happens to v
I In both cases, noise traders trade
I buy and sell u quantities in expectation
I Only if innovations happen to v, informed traders trade
I trade with intensity µ
I always in the right trading direction

I This setting incorporates event (or model) uncertainty

30 / 36
Information Measure: Event Uncertainty

31 / 36
Information Measure: Event Uncertainty

I Assuming Poisson arrival,1 we can write the joint probability of

having buy (b) and sell (s) as function of parameters

Pr(b, s) = f (α, µ, u)

I think of is as the number of b and s in a day

I Using order flow data, estimate α, µ, and u by maximum

likelihood
I Use estimated values to evaluate information frictions (and
spread)

1 Beyond the scope of the class. Refer to Ch 6.2 of Hasbrouck (2007)

32 / 36
Information Measure: PIN

I The expected total order arrival intensity is

2u + αµ

I Probability of informed trading (PIN):

αµ
PI N =
2u + αµ

I Prob that a randomly chosen trader on a randomly chosen day is

informed

I Ex., applying to the simplified GM model, the spread for the first
trade should be
A−B
S= = PI N × σ
2

33 / 36
PIN Analyses
I Easley, et al., (2002): estimated PIN for all NYSE common stocks

34 / 36
PIN Analyses

I Easley, et al., (2002): estimated PIN for all NYSE common stocks
I median PIN = 19%
I difference across stocks is not large

I PIN is negatively correlated with firm size (ρ = −0.58)

I PIN is positively correlated in cross-section with
1. stock return volatility (ρ = 0.24)
2. bid-ask spread (ρ = 0.35)

I PIN is not correlated with volume

35 / 36
PIN Analyses (cont’d)

I Some other observations of PIN:

I PIN is higher in electronic trading markets than in non-anonymous
venues (like floor trading)
I Experiment with German Stock Exchange
I PIN is higher for stocks with less analyst coverage
I well known that spread widens when analysts stop following a stock

I Due to DeFi and blockchain, a new identification method appears;

how will it change?

36 / 36