Optimal Execution

&
Algorithmic Trading

Jan Oblój
jan.obloj@maths.ox.ac.uk

NUS Students’ Training Week at MCFG

Mathematical Institute, University of Oxford

17 – 21 June 2019

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 1 / 74
 Outline Reading
Modelling in Quantitative Finance There is a wealth of ongoing research and growing body of publications.
Brief history of modelling in QF For market impact modelling, a good start are two survey papers:
LOBs Impact – recall • Lehalle,“Market Microstructure Knowledge Needed for Controlling an Intra-Day Trading
Process”
Market Frictions • Gatheral and Schied, “Dynamical models of market impact and algorithms for order
execution”
Price Impact Models and Optimal Execution
both in 2013 Handook on Systemic Risk (ed. Fouque and Langsam) and on arXiv.
The modelling setup For market microstructure, I suggest two review papers and four books:
Almgren–Chriss models • Chakraborti et al, “Econophysics review” (parts I and II), in Quantitative Finance, 2011
• “How markets slowly digest changes in supply and demand”, Bouchaud et al (2009)
Transient Price Impact Models • O’Hara, Market Microstructure Theory, 1995
Obizhaeva–Wang type models • Hasbrouck, Empirical Market Microstructure: The Institutions, Economics, and
Non-robustness w.r.t. decay kernel Econometrics of Securities Trading, 2006
Regularity of market models • Lehalle and Laruelle, Market Microstructure in Practice, 2014.
• Cartea, Jaimungal and Penalva, Algorithmic and High-Frequency Trading, 2015.
Predatory trading and HF hot-potatos

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 2 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 3 / 74

Modelling in Quantitative Finance

Outline
Modelling in Quantitative Finance
Brief history of modelling in QF
LOBs Impact – recall
Market Frictions

Price Impact Models and Optimal Execution

The modelling setup
Almgren–Chriss models

Transient Price Impact Models

Obizhaeva–Wang type models
Non-robustness w.r.t. decay kernel
Regularity of market models

Predatory trading and HF hot-potatos

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 4 / 74
 Modelling in Quantitative Finance Brief history of modelling in QF Modelling in Quantitative Finance Brief history of modelling in QF

Models in Quantitative Finance Brief history of modelling in QF

• Fair price (fundamental economics)

• “All models are wrong but some are useful” (G. Box ’78)
model: fundamentals
• Models need to be tailored to
• Option pricing & optimal investment
• the available inputs
• the intended outputs model: the underlying price process (exogenous)
Samuelson ’65, B&S and Merton ’73
• Models need to
• conform to stylised facts • Further option pricing: Exotics or FI options
• produce reasonably useful and robust outputs model: a high- or ∞- dimensional system of underlyings
• avoid creating arbitrage opportunities e.g.: HJM ’92 and LMM ’97 in Fixed Income; Market models of
Schweizer & Wissel ’08, Carmona & Nadtochiy ’09

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 5 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 6 / 74

Modelling in Quantitative Finance Brief history of modelling in QF

Brief history of modelling in QF – cont.

• Optimal execution of planned trades

model: impact of trades on price dynamics or
model: supply & demand dynamics
Bertsimas & Lo ’98, Almgren & Chriss ’00;
Obizhaeva & Wang ’13, Alfonsi et al. ’08
• Price formation via market microstructure
model: LOB dynamics (zero intelligence)
model: Agent trades (agent based)
Cont et al. ’10, Smith et al. ’03, Farmer et al. ’05;
Kyle ’85, O’Hara ’95
... and MANY more references...

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 7 / 74
 Modelling in Quantitative Finance LOBs Impact – recall Modelling in Quantitative Finance LOBs Impact – recall

Recall: LOBs Impact – pros Recall: LOBs Impact – cons

And some negative:
The shift from traditional markets to electronic LOB driven markets had • Technological armsrace
many consequences. Some positive: • Little human oversight
• competition leading to lower fees and smaller tick sized • Predatory trading
• more information available This led to the infamous Flash Crash of May 6, 2010 when Dow Jones IA
• democratised trading process (DJIA) dived almost 1000 points (just to recover in minutes).
• choice of patient (limit) or impatient (market) orders available to
everyone
• computerised/algorithmic trading possible
• high frequency trading possible
• HFT ≈ duration of order of seconds, reaction within milliseconds
• accounts for 60 − 75% of traded volume
• extra provision of liquidity market efficiency

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 8 / 74 What
c Jan Obdo you think
lój, University caused
of Oxford Price it?
Impact Models and Market Microstructure 17 – 21 June 2019 9 / 74

• A mutual fund activated a program to sell 75, 000 E-Mini S&P 500
contracts (≈ 4.1 billion USD) using VWAP algorithm at 9%
Modelling in Quantitative Finance LOBs Impact – recall • HFT began to quickly buy and resell these contracts to each other
generating more volume: between 2:45:14 and 2:45:27, HFT traded
Flash Crash of May 6th, 2010 27, 000 contracts (about 49% of total volume) while buying only 200
contracts net.
• This led the original program to rapidly sell the whole position

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 10 / 74
 Modelling in Quantitative Finance Market Frictions Modelling in Quantitative Finance Market Frictions

Frictionless modelling setting Market Frictions

The classical modelling framework in mathematical finance, like the one

postulated by Black and Scholes ’73, assumes infinite liquidity:
• asset traded at uniquely given and known prices
• buying and selling in arbitrary quantities possible
• trading at no cost possible
• trading has no impact on the price
This is unrealistic and unsatisfactory: in reality we have market frictions.

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 11 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 12 / 74

Modelling in Quantitative Finance Market Frictions

Market frictions – cont.

Many frictions either part of the game (opportunity cost) or well-defined

(taxes). For many traders other frictions satisfactory summarised in
• proportional transaction costs: pay St for trading one unit of St .
However this is not acceptable for
• large trades (relative to volume & time horizon)
• frequent trading (relative to liquidity)
which require understanding of
• liquidity provision and
• price formation.

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 13 / 74
 Modelling in Quantitative Finance Market Frictions Modelling in Quantitative Finance Market Frictions

Chart 1
and exit (Kyle
Aspects of liquidity at no cost,
’85) transparent information). The Aspects ofAspects of liquidity
market liquidity (Kyle ’85)
degree of liquidity of a market is traditionally
assessed on the basis of three essential criteria: Price

• the tightness of the bid-ask spread, which measures

• Tightness (Breadth): measures how wide the bid-ask is, i.e. measures
the cost of a reversal of position at short notice for Depth
the cost of a position reversal at a short notice for a standard amount
a standard amount, Ask price Ap Resilience
• Market depth: corresponds to the volume which may be
bought/sold without immediately affecting
• market the price
depth, which corresponds to the volume Breadth
• Market resilience: describesofthetransactions
speed at whichthat may
prices be immediately
revert to executed Depth
previous level (equilibrium) after a random shock in the order flow
without slippage of best limit prices,
Resilience Bp Bid price
• Time delay: measures the delay between processing and executing
an order • market resilience, i.e. the speed with which prices
revert to their equilibrium level following a random
Quantities Quantities
shock in the transaction flow.
A 0 A'
Sale Purchase
The first aspect is a direct measure of transaction
Source: Bervas ’06
c Jan Oblój, University of Oxford
costs (excluding other
Price Impact Models and Market Microstructure
operational
17 – 21 June 2019 14 / 74
costs such as c Jan Oblój,
Note : The
University bid price
of Oxford BpImpact
Price andModels
the and
askMarket
price Ap are defined
Microstructure for
17 – the 2019
21 June 15 / 74
brokerage commissions and clearing and settlement standard amounts OA and OA’. The Bp-Ap spread represents the
fees). The last two indicate the market’s ability to “breadth” of the market. The amounts OA and OA’ are those that
may be traded without price slippage: they reflect market “depth”.
absorb
Modelling in Quantitative Finance Marketsignifi
Frictions cant volumes without adverse effects
Beyond points A and A’, one sees the negative impact of large-value
on prices. The rest of the article will focus mainly transactions on the execution price. Resilience refers to the time aspect
Summary on marketso farbreadth and depth insofar as it will pay of liquidity and indicates how quickly prices adjust to their equilibrium
more attention to the costs of immediacy than to value following a shock in transaction flow.

• Models are build taking into how account long it takes

available prices
inputs to return to equilibrium
and desirable
outputs (see Chart 1). investors that have private information regarding
• In QF models postulate exogenous dynamics for different underlyings the real value of the asset). In a quote-driven market,
depending on what is traded The bid price
and what one wants is the highest price that the market
to price the quoted spread2 corresponds to the difference
maker is willing to pay at a given time to acquire
• Traditional models assume a frictionless setting with ∞ liquidity
between the best bid price and the best ask price
a specific amount of assets. Symmetrically, the offered by market makers, whilst in an order-
• In practice this fails. A lot can be accounted for using proportional
ask price is the lowest price at which the market driven market, what is important is the difference
transactions costs. maker is willing to sell a given amount of assets. between the best limit order book prices.3 However,
• Large and/or frequent tradingThe gap between the bid price and the ask price
requires modelling of liquidity and/or the spread quoted in the markets is not generally
price impact. (the bid-ask spread) compensates the market maker an exact reflection of transaction costs (for a
• Electronic markets operate without for thedesignated
immediacy of execution
market maker. that it offers to its buy/sell sequence) because certain transactions
counterparties.
• Instead, the Limit Order Book (LOB) holds all active buy and sell The spread measures the cost of a may be traded not at the bid or the ask price but at
orders sell/buy or buy/sell sequence over a short period prices located within this spread, or even outside
(two-way transaction); only the half-spread should this spread, even for standard amounts.4 In addition,
therefore be attributed to a single transaction (sale or the spread is a measure of the liquidity available
c Jan Oblój, University of Oxford
purchase) if one considers
Price Impact Models and Market Microstructure
that the
17 – 21 June 2019
mid-price is the
16 / 74
at a given time. With a view to risk measurement
 Price Impact Models and Optimal Execution Price Impact Models and Optimal Execution The modelling setup

Outline Price impact modelling

Modelling in Quantitative Finance
Brief history of modelling in QF We saw that large and/or frequent trades may affect the price. We may
LOBs Impact – recall need to split and spread large orders in practice. To answer how to do it
Market Frictions we need to understand:
• how to model/quantify the impact of trading on the price?
Price Impact Models and Optimal Execution
• what are the desirable/undesirable properties of such models?
The modelling setup
Almgren–Chriss models • how to compute optimal execution trading strategies?
There are two natural approaches to model price impact:
Transient Price Impact Models
I: postulate fair price dynamics and the price impact of trading
Obizhaeva–Wang type models
Non-robustness w.r.t. decay kernel II: be serious about modelling Market Microstructure, i.e. model supply
Regularity of market models and demand and their interaction.
We focus first on I. Then we use the LOB discussion to tackle II.
Predatory trading and HF hot-potatos

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 17 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 18 / 74

Price Impact Models and Optimal Execution The modelling setup

Trade execution setup

Goal: buy/sell x0 shares by time T .
Trade execution strategy:
• X = (Xt )t≤T , where Xt is the number of shares held at time t
• The initial position X0 = x0 is positive for a sell strategy and negative
for a buy strategy
• The final condition XT = 0 indicates the position is liquidated at T
• The path will be monotone for a pure buy or pure sell strategy. In
general it is of finite variation.
We think of T as around 5 − 10, and up to 30, minutes.
For now, we are ignoring problems from higher(+) or lower(-) levels:
+ How a large desired trade position is split into chunks allocated their
time horizons.
- What orders (market vs limit) are used and to which venues these are
routed.
c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 19 / 74
 Price Impact Models and Optimal Execution The modelling setup Price Impact Models and Optimal Execution The modelling setup

Price impact model Revenues and costs

Suppose Xt is differentiable in time and StX depends continuously on X ,
Price impact model quantifies the feedback effect of trading strategy X on then at time t, the infinitesimal amount of −dXt shares is sold at price
the asset price. A typical setup is: StX . Thus
• Exogenously specified price process S 0 = (St0 : t ≤ T ) for fair
Z T
(unaffected) price dynamics.
revenues from strategy X are R(X ) = − StX dXt
S is a semimartingale (usually a martingale) on a filtered probability 0
space (Ω, F, (F t ), P) and we assume X is predictible
(when X is not absolutely continuous adjustments are necessary)
• Given X , a model prescribes S X the price process realised when
implementing trading strategy X . Objective: Maximise some performance functional of R(X ).
For example:
• Typically, a buy strategy increases the prices and a sell strategy
• maximise the expected value E[R(X )]
decreases the prices: if X 0 (t) ≥ 0 for all t ≤ T then StX ≥ St0 , t ≤ T .
However this is not necessarily true for a fixed t since StX may be • maximise a mean-variance criterion E[R(X )] − λ var(R(X ))
affected by all of (Xu : u ≤ t). • maximise the expected utility E[U(R(X ))]
• ...

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 20 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 21 / 74

Price Impact Models and Optimal Execution The modelling setup

Revenues and costs – cont.

Alternatively: Minimise functional of implementation shortfall (i.e. cost
of liquidation), which is the difference between the book value X0 S00 and
the revenues (or the capture):
liquidation cost of X is C(X ) = X0 S00 − RT (X ).
If we write StX = St0 + ItX then
Z T Z T Z T
X 0
R(X ) = − St dXt = − St dXt − ItX dXt
0 0 0
Z T Z T
0 0
= S0 X0 + Xt dSt − ItX Ẋt dt
0 0
| {z } | {z }
=− C vol (X ) =C exec (X )

The total liquidation cost C(X ) has two components:

• C vol expresses the volatility risk of trading over time instead of
instantly
• C exec expresses the effect of price impact
c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 22 / 74
 Price Impact Models and Optimal Execution Almgren–Chriss models Price Impact Models and Optimal Execution Almgren–Chriss models

Almgren–Chriss type price impact A–C model with linear price impact
The unaffected price follows a Brownian motion:

St0 = S00 + σWt . In the special case of linear impacts: g (x) = γx and h(x) = ηx
Z t
Then, the price impact has two components: X 0
St = St + γ dXs + η Ẋt = St0 + γ(Xt − X0 ) + η Ẋt .
• permanent impact: 0t g (Ẋs )ds
R
0 | {z }
=ItX
• temporary impact: h(Ẋt )
for nondecreasing functions g , h : R → R and Ẋt = dX dt the trading speed.
t The revenues are then given by
The affected price is given by Z T Z T Z T
Z t R(X ) = − StX dXt = S00 X0 + Xt dSt0 − ItX Ẋt dt
X 0
St = St + g (Ẋs )ds + h(Ẋt ). 0 0 0
0
Z T Z T
γ
= S00 x0 +σ Xt dWt − x02 − η Ẋt2 dt,
In the special case of linear impacts: g (x) = γx and h(x) = ηx 0 2 0
Z t
X 0 since XT = 0.
St = St + γ dXs + η Ẋt = St0 + γ(Xt − X0 ) + η Ẋt .
0

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 23 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 24 / 74

Price Impact Models and Optimal Execution Almgren–Chriss models

A–C model with linear price impact (cont.)

Assuming X is bounded, the expected revenues are
Z T 
0 γ 2 2
E[R(X )] = S0 x0 − x0 − η E Ẋt dt .
2 0

The last term is an integral w.r.t. P(dω) ⊗ dt of the square of Ẋt (ω). It
follows that it is minimised, and hence E[R(X )] is maximised, by the
strategy
x0
Ẋt∗ = −
T
which sells (or buys) the shares at constant speed (to see this simply apply
Jensen’s inequality). In particular the solution is independent of the
volatility! (Bertsimas & Lo ’98)
The resulting expected liquidation cost of x0 shares is
γ 
E[C(X )] = + η x02
2
quadratic in number of shares and independent of volatility σ.
c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 25 / 74
 Price Impact Models and Optimal Execution Almgren–Chriss models Price Impact Models and Optimal Execution Almgren–Chriss models

A–C model so far – summary A–C model with mean-variance criterion

Proposition So far we only looked at expected revenues. Almgren and Chriss ’00
In the Almgren–Chriss price impact model with linear permanent impact, propose to consider
g (x) = γx, and xh(x) convex, for any given x0 ∈ R the strategy
max E[R(X )] subject to var(R(X )) ≤ v∗
X0 (T − t) X
Xt∗ = , t ≤ T,
T which, introducing a Langrange multiplier, turns into an unconstrained
maximises the expected revenues E[R(X )] in the class of all adapted and problem
bounded trade execution strategies X . max (E[R(X )] − λ var(R(X ))) .
X
The strategy X ∗ spreads the execution evenly over the time horizon This is a hard problem. However assuming X is deterministic it turns into
t ∈ [0, T ]. It is often referred to as the time-weighted average price
Z T 2
strategy or TWAP. When the time is relative and t corresponds to traded
  
0 γ 2 λσ 2 2
volume the X ∗ is called volume-weighted average price strategy or VWAP. max x0 S0 − x0 − X + η Ẋt dt
X 2 0 2 t
Both are used as industry benchmarks.
Almgren et al. ’05 argued these assumptions are consistent with empirical which can be solved explicitly as a standard variational calculus problem.
observations and suggested xh(x) ≈ |x|1.6 .
c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 26 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 27 / 74

Price Impact Models and Optimal Execution Almgren–Chriss models

A–C model with mean-variance criterion

Indeed, the problem is equivalent to
Z T 2 
λσ 2 0 2
min X (t) + ηX (t) dt
X 0 2
Setting the first variation to zero:
Z T
g (t)λσ 2 X (t) + 2g 0 (t)ηX 0 (t) dt, ∀g ∈ C 1 : g (0) = g (T ) = 0.


0=
0
Integrating by parts:
Z T
g (t) λσ 2 X (t) − 2ηX 00 (t) dt, ∀g ∈ C 1 : g (0) = g (T ) = 0


0=
0
which gives the Euler-Lagrange equation
λσ 2
X 00 (t) = X (t), s.t. X (0) = x0 , X (T ) = 0.
2η
Solving the ODE we obtain
c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 28 / 74
 Price Impact Models and Optimal Execution Almgren–Chriss models Price Impact Models and Optimal Execution Almgren–Chriss models

A–C model with mean-variance criterion A–C model with mean-variance criterion
The solution is given by The solution is given by
s s
sinh(κ(T − t)) λσ 2 sinh(κ(T − t)) λσ 2
Xt∗ = x0 for κ = . Xt∗ = x0 for κ = .
sinh κT 2η sinh κT 2η

1 × 106 1 × 106

800 000 800 000

600 000 600 000

400 000 400 000

200 000 200 000

1 2 3 4 5 1 2 3 4 5

6 6
Optimal liquidation strategy of 10 shares over 5 days under 30% annual vol and Optimal liquidation strategy of 10 shares over 5 days under 30% annual vol and
impact 1% of daily volume = bid-ask. Moderate λ. impact 1% of daily volume = bid-ask. High λ.
c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 29 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 30 / 74

Price Impact Models and Optimal Execution Almgren–Chriss models

A–C model with other criteria

Mean-variance is not amenable to dynamic programming and leads to

time-inconsistent strategies. In analogy to optimal investment, other
criteria are natural:
• Maximise expected utility: maxX E[U(R(X ))]
The problem can be reformulated as a stochastic control problem with
non-standard (finite fuel) constraint: X0 = x0 and XT = 0. Leads to
an HJB equation. Solution known for U(x) = − exp(−λx) ... the
same as for mean-variance! (Schied, Schöneborn & Tehranchi ’10).
• Maximise  Z  T
E R(X ) − λ Xt StX dt
0
Gatheral & Schied ’11

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 31 / 74
 Price Impact Models and Optimal Execution Almgren–Chriss models Price Impact Models and Optimal Execution Almgren–Chriss models

Criticism of A–C setting Summary of A–Ch-type market impact modelling

• Price process can go negative; impact additive & in absolute terms.
Bertsimas & Lo ’98 suggest • Revenues from a large sell/buy order may depend crucially on its
Z t execution
σ2
  
X 0
St = St exp 0 0
g (Ẋs )ds + h(Ẋt ) , St = S0 exp σWt − t • The optimal execution strategy in turn may depend crucially on the
0 2 criterion
but computing optimal strategies more involved. • Almgren–Chriss models involve permament and temporary impact of
• Price impact simplistic, in reality transient effect, see Moro et al. ’09 trades on prices
(cf. resilience) • Under linear impacts and maximising revenues, it is optimal to sell at
• Computed optimal strategies are deterministic and do not react to a constant speed
price changes • Under linear impacts and among deterministic strategies, optimising
• No modelling of feedback effects between the seller and the market mean-variance criterion, it is optimal to use a specific convex
(e.g. Flash Crash 06/05/10) programme.

=⇒ Need to understand price formation better!

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 32 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 33 / 74

Transient Price Impact Models

Outline
Modelling in Quantitative Finance
Brief history of modelling in QF
LOBs Impact – recall
Market Frictions

Price Impact Models and Optimal Execution

The modelling setup
Almgren–Chriss models

Transient Price Impact Models

Obizhaeva–Wang type models
Non-robustness w.r.t. decay kernel
Regularity of market models

Predatory trading and HF hot-potatos

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 34 / 74
 Transient Price Impact Models Obizhaeva–Wang type models Transient Price Impact Models Obizhaeva–Wang type models

Types of price impact Modelling transient price impact

So far we have modelled: Idea: model transient price impact by:

• permanent price impact • stochastic dynamics of LOB
• temporary price impact e.g. constant depth λ, model only bid Bt & ask At
• a buy (market) order eats into the ask side of the book
In reality, transactions interact with the LOB. Market orders will eat into
the book but new liquidity will then come as markets are resilient. a buy order of ∆Xt > 0 moves ask At+ = At + ∆Xt /λ
• book then reverts back at some speed
We need to model
according to a decay kernel G (delay), e.g. e−ρt , (1 + t)−α
• transient price impact
Obizhaeva & Wang ’13, Alfonsi et al. ’08, Gatheral ’10, Gatheral et al. ’12...

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 35 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 36 / 74

Transient Price Impact Models Obizhaeva–Wang type models

Simple transient price impact (Obizhaeva & Wang ’13)

• Assume no bid-ask spread, St0 = Bt = At is a martingale
• Constant book depth of λ = 1/G (0)
• A discrete order Xt+ − Xt =: ∆Xt moves price
X
St+ = StX + ∆Xt G (0)

and is executed at cost of (= - expected revenue of)

Z X
St+
1 1  X 2  G (0)
vdv = (St+ ) − (StX )2 = (∆Xt )2 +∆Xt StX .
G (0) StX 2G (0) 2

• The market is resilient and trade impact wanes away. So that

X
StX = St0 + G (t − s)∆Xs
s<t:∆Xs >0

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 37 / 74
 Transient Price Impact Models Obizhaeva–Wang type models Transient Price Impact Models Obizhaeva–Wang type models

Simple transient price impact – cont. Simple transient price impact – cont.
• Assume now trading is only possible at some give time points: n
X Z t Z t
0 = t0 < t1 < . . . < tn = T , X0 given, XT = 0 and S00 X0 + St0i ∆i = S00 X0 + St0 dXt =− Xt− dSt0
X
i=0 0 0
Xt = X0 + ∆i , where ∆i := Xti + − Xti
i:ti <t which has zero expectation (assuming ∆i bounded). Further,
 
• The mid-price resulting from strategy X is n
 G (0) ∆2i + ∆i
X X
X G (ti − tj )∆j 
StX = St0 + G (t − ti )∆i 2
i=0 j<i
i:ti <t X G (0) XX
• The total cost of executing X is = ∆2i + G (ti − tj )∆i ∆j
2
n   i i j<i
X G (0) 1 XX
C(X ) = S00 X0 − R(X ) = S00 X0 + ∆2i + ∆i StX = G (|ti − tj |)∆i ∆j
2 2
i=0 i j
 
n n
X X
 G (0) ∆2i + ∆i
X In consequence, the total expected cost of liquidation following X is
= S00 X0 + St0i ∆i + G (ti − tj )∆j 
2 1 XX
i=0 i=0 j<i E[C(X )] = G (|ti − tj |) E [∆i ∆j ]
2
i j
c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 38 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 39 / 74

Transient Price Impact Models Obizhaeva–Wang type models

Simple transient price impact – solution

It is then enough to look for X among deterministic strategies:

XX
minimise G (|ti − tj |)∆i ∆j over ∆ ∈ Rn+1 : ∆T 1 = −x0
i j

Rk: value invariant under ∆ → −∆ =⇒ Optimal Buy = - Optimal Sell.

If G is strictly positive definite then the optimal solution ∆∗ is

∆∗ = const · Γ−1 1, where Γij = G (|ti − tj |).

T
Let us take equidistant steps: ti+1 − ti = N and look at different
examples of G .

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 40 / 74
 Transient Price Impact Models Obizhaeva–Wang type models Transient Price Impact Models Obizhaeva–Wang type models

Optimal strategy – examples Optimal strategy – examples

Optimal ∆∗i for t ∈ [0, 1], N = 20, X0 = −100 and four decay kernels: Optimal ∆∗i for t ∈ [0, 1], N = 100, X0 = −100 and four decay kernels:
1 1 1 1 1
G1 (t) = e−5t , G2 (t) = (0.5 − 2.7t)+ , G3 (t) = , G4 (t) = . G1 (t) = (0.5−2.7t)+ , G2 (t) = , G3 (t) = G4 (t) = .
(1 + 10t)2 1 + (10t)2 (1 + 5t)2 1 + (10t)2 1 + (7t)2
14 18 5
x 10
14 1.5

16
12
12 1
14

10
12 10
0.5

8 10
8
0
8
6
6
6
−0.5
4
4
4

2 −1
2 2

20
0
−5 0 5 10 15 20 25

9
0
−5 0 5 10 15 20 25
Which one is which? 0
−20 0 20 40 60 80 100 120
−1.5
−20 0 20 40 60 80 100 120
Which one is which?
8
x 10 6
1.5
8

15
7 1 5

10 0.5 4

4 0 3
5

3
−0.5 2

2
0

1 −1 1

−5 0
−5 0 5 10 15 20 25 −5 0 5 10 15 20 25
−1.5 0
c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 41 / 74 −20
c Jan Oblój, University of Oxford
0 20 40 60 80 100
Price
120 −20
Impact
0
Models
20
and Market
40 60 80
Microstructure
100 120
17 – 21 June 2019 42 / 74

Transient Price Impact Models Non-robustness w.r.t. decay kernel

Non-robustness w.r.t. decay kernel

The optimal ∆∗i for t ∈ [0, 1], N = 100, X0 = −100 and three decay kernels:

1 1 1
G2 (t) = 2
, G3 (t) = 2
G4 (t) = .
(1 + 5t) 1 + (10t) 1 + (7t)2

1.0

0.8

0.6

0.4

0.2

0.0 0.2 0.4 0.6 0.8 1.0

differ dramatically...
c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 43 / 74
 Transient Price Impact Models Non-robustness w.r.t. decay kernel Transient Price Impact Models Non-robustness w.r.t. decay kernel

Notion of “price manipulation strategy” Transient price impact with arbitrary strategies
With discrete X , the impacted price process was
X Z
X 0 0
We saw that very similar decay functions may lead to drastically different St = St + G (t − ti )∆i = St + G (t − s)dXs
i:ti <t s<t
optimal portfolios, including round-trip-taking trading. Clearly requires
further studies. and the last term extends to arbitrary X (predictable, left-continuous, of
bounded variation). The revenues of a continuous strategy are given as
Definition previously
A round trip strategy X , X0 = XT = 0 with strictly negative expected cost Z T Z T Z TZ
E[C(X )] < 0 is called a price manipulation strategy. − X
St dXt = − 0
St dXt − G (t − s)dXs dXt .
Note that this is not the usual arbitrage since profit is not a.s. but in 0 0 0 s<t
expectation. However in some models rescaling and repeating price In the case of discrete X we had
manipulation leads to (weak) arbitrage. Xn
1 XX
− St0i ∆i − G (|ti − tj |)∆i ∆j
We first extend our previous analysis to arbitrary strategies X . 2
i=0 i j
Z T Z Z
0 1
=− St dXt − G (|t − s|)dXs dXt
0 2
c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 44 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 45 / 74

Transient Price Impact Models Non-robustness w.r.t. decay kernel

Transient price impact with arbitrary strategies

Combining, the execution cost of X are

Z T Z T Z T
1
C(X ) = S00 X0 − R(X ) = Xt− dSt0 + G (|t − s|)dXs dXt .
0 2 0 0

composed of volatility risk and price impact cost

Z T Z T
exec 1
C (X ) = G (|t − s|)dXs dXt .
2 0 0

Price manipulation ⇐⇒ E[C exec (X )] < 0.

Let’s start with understanding when C exec (X ) ≥ 0 a.s.

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 46 / 74
 Transient Price Impact Models Non-robustness w.r.t. decay kernel Transient Price Impact Models Non-robustness w.r.t. decay kernel

Bochner’s theorem and positive costs Is absence of price manipulation enough?

Proposition
We have C exec (X ) ≥ 0 for all strategies X iff G is positive definite, i.e. can be
represented as the Fourier transform of a positive finite Borel measure µ on R. We have
Further, if G is strictly positive definite (µ is not discrete) then C exec (X ) > 0 for
positive definite G =⇒ no price manipulation strategy.
all nonzero X .
We may also formalise the case of deterministic discrete strategies. Is this enough? Take
2

Proposition (Gatheral, Schied and Slynko ’12) G (t) = e−t

Suppose G is positive definite. Then among deterministic strategies trading at which, up to scaling, is its own Fourier transform and hence positive
given times (ti ), an optimal one X ∗ satisfies a generalised Freedholm integral definite.
equation
Let’s look at the optimal strategy for T = 10, X0 = −100 and vary N.
Z
G (|ti − s|)dXs∗ = λ, i = 0, 1, . . . , N

for some constant λ.

Rk.: We wrote this equation as Γ∆ = const · 1 before.

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 47 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 48 / 74

Transient Price Impact Models Non-robustness w.r.t. decay kernel

Optimal trading with Gaussian decay G (t) = exp(−t 2 ),

T = 10, X0 = −100, N = 10
14

12

10

0
0 1 2 3 4 5 6 7 8 9 10

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 49 / 74
 Transient Price Impact Models Non-robustness w.r.t. decay kernel Transient Price Impact Models Non-robustness w.r.t. decay kernel

Optimal trading with Gaussian decay G (t) = exp(−t 2 ), Optimal trading with Gaussian decay G (t) = exp(−t 2 ),
T = 10, X0 = −100, N = 15 T = 10, X0 = −100, N = 20
16 40

14
30

12

20
10

8 10

6
0

4
−10

−20
0

−2 −30
−2 0 2 4 6 8 10 12 14 16 −5 0 5 10 15 20 25

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 50 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 51 / 74

Transient Price Impact Models Non-robustness w.r.t. decay kernel

Optimal trading with Gaussian decay G (t) = exp(−t 2 ),

T = 10, X0 = −100, N = 25
200

150

100

50

−50

−100

−150

−200
−5 0 5 10 15 20 25 30

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 52 / 74
 Transient Price Impact Models Non-robustness w.r.t. decay kernel Transient Price Impact Models Non-robustness w.r.t. decay kernel

Optimal trading with Gaussian decay G (t) = exp(−t 2 ), Optimal trading with Gaussian decay G (t) = exp(−t 2 ),
T = 10, X0 = −100, N = 37 T = 10, X0 = −100, N = 38
5 6
x 10 x 10
3 2

1.5
2

1
0.5

0 0

−0.5
−1

−1

−2
−1.5

−3 −2
−5 0 5 10 15 20 25 30 35 40 −5 0 5 10 15 20 25 30 35 40

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 53 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 54 / 74

Transient Price Impact Models Non-robustness w.r.t. decay kernel

Optimal trading with Gaussian decay G (t) = exp(−t 2 ),

T = 10, X0 = −100, N = 100
8
x 10
4

−1

−2

−3

−4

−5
−20 0 20 40 60 80 100 120

Clearly excluding price manipulation strategies is not enough...

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 55 / 74
 Transient Price Impact Models Regularity of market models Transient Price Impact Models Regularity of market models

Price manipulation strategies Regularity of Almgren–Chriss type models

Definition Recall that in A-CH framework, the impacted price is

A market model admits price manipulation if there exists a round trip Z t
strategy X , X0 = XT = 0 with strictly positive expected revenues X 0
St = St + g (Ẋs )ds + h(Ẋt ).
E[R(X )] > 0. 0

Definition Proposition (Huberman & Stanzl ’04, Gatheral ’10)

We say that a market impact model admits transaction–triggered price
If the model above does NOT admit price manipulation for all T > 0 then
manipulation if the expected revenues of a sell (resp. buy) program can be
g (x) = γx for some γ ≥ 0.
increased by intermediate buy (resp. sell) orders.
Further, if g is linear and x → xh(x) is convex than the model does NOT
Remark: in a sensible model (i.e. if buying increases prices and selling admit transaction–triggered price manipulation.
decreases prices) absence of transaction–triggered price manipulation
Rk: the second part is clear since in this setting the optimal X ∗ is linear.
implies absence of the usual price manipulation.

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 56 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 57 / 74

Transient Price Impact Models Regularity of market models

Regularity of Obizhaeva–Wang type models

Proposition (Alfonsi, Schied & Slynko ’12)

A transient price impact model with decay kernel G s.t.

G (0) − G (s) < G (t) − G (t + s), for some s 6= t,

admits transaction-triggered price manipulation trading at {0, s, t + s}.

In particular, it is enough that G is NOT convex for small t.

Proposition (Alfonsi et al. ’12, Gatheral et al. ’12)

A transient price impact model with convex, decreasing, non-negative
decay kernel G admits a unique optimal X ∗ which is monotone in time. In
particular the setup does NOT admit transaction-triggered price
manipulation.

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 58 / 74
 Transient Price Impact Models Regularity of market models Transient Price Impact Models Regularity of market models

Other developments Summary of transient market impact models

• Non-linear transient price impact models: the book has varying depth • Transient price impact models take into account the interaction of
according to a given shape f , see Alfonsi & Schied ’10 orders with the LOB and market resilience
• A combination of impacts, e.g. Gatheral ’10 • Under constant LOB depth, discrete trading at (ti ) and maximising
Z t expected revenues the optimal strategy explicit for many impact
StX = St0 + h(−Ẋt )G (t − s)ds decay kernels G
0
• More generally the problem quickly becomes very hard...
• Stochastic models of LOB where the shape f is a stochastic process • Even in simple setting, the optimal strategies may often involve round
in space of curves and/or stochastic resilience, see Alfonsi & Infante trips. Solution is non-robust with respect to G .
Acevedo ’12, Klöck ’12, Fruth, Schöneborn & Urusov ’11, Müller &
• Possible to study, and provide sufficient conditions for, the absence of
Keller-Ressel ’15.
price-triggered manipulation strategies.
• ...

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 59 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 60 / 74

Predatory trading and HF hot-potatos

Outline
Modelling in Quantitative Finance
Brief history of modelling in QF
LOBs Impact – recall
Market Frictions

Price Impact Models and Optimal Execution

The modelling setup
Almgren–Chriss models

Transient Price Impact Models

Obizhaeva–Wang type models
Non-robustness w.r.t. decay kernel
Regularity of market models

Predatory trading and HF hot-potatos

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 61 / 74
 Predatory trading and HF hot-potatos Predatory trading and HF hot-potatos

Multi-agent frameworks Predatory Trading

Large Trader facing a forced liquidation

+
• In reality many agents interact in a market. other (HF) traders aware of this fact
• Mathematically best modelled as game. When number of players ⇓
n → ∞, sometimes possible to analyse as a mean field game. Predatory Trading
• Interesting as it allows to study
• Interaction of one large player with n small players (e.g. predatory Examples of “targets”:
trading) • Index-replicating funds at rebalancing dates
• Global market implications of interactions between small players • Institutional investors subject to regulatory constraints (e.g. when an
• Properties of markets which facilitate different phenomena
instrument is downgraded)
• Traders using portfolio insurance or stop-loss strategies
• Hedge funds close to a margin call
• Recalled short-seller

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 62 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 63 / 74

Predatory trading and HF hot-potatos

Predatory Trading

“... if lenders know that a hedge fund needs to sell something quickly, they
will sell the same asset – driving the price down even faster. Goldman,
Sachs & Co. and other counterparties to LTCM did exactly that in 1998.”
Business Week, 26 Feb 2001

“When you smell blood in the water, you become a shark ... when you
know that one of your number is in trouble ... you try to figure out what
he owns and you start shorting those stocks ... ”
Cramer, 2002

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 64 / 74
 Predatory trading and HF hot-potatos Predatory trading and HF hot-potatos

Predatory Trading – mechanisms One-period game model with A–Ch price impact
When a need of a large trader (prey) to liquidate is recognised, the • n + 1 players with portfolios X0 (t), . . . , Xn (t), t ∈ [0, T ], assumed
strategic traders (predators) might cont. diff. in time
• first trader in the same direction • one prey (seller): X0 (0) = x0 > 0, X0 (T ) = 0
• withdraw liquidity instead of providing it • n predators: Xi (0) = Xi (T ) = 0, i = 1, . . . , n
• market impact is greater leading to price overshooting • and the above is common knowledge
• may further enforce distressed trader’s need to liquidate
• players are risk-neutral and maximise their expected profit
• then reverse direction to profit from the overshoot
Z T 
• closing the roundtrip at a profit. i
R (X ) = − E St dXi (t)
However when strategic traders have a longer horizon than the liquidation, 0
their behaviour may depend on market characteristics: • one risk-free and one risky asset, continuous trading, Almgren–Chriss
• could act as predators as above large trader tries to keep linear price impact model
intentions hidden (stealth trading)
n n
• could act as liquidity providers large trader announces intentions S(t) = S(0) + σWt + γ
X
(Xi (t) − Xi (0)) + η
X
Ẋi (t)
(sunshine trading) i=1 i=1
see Brunnermeier & Pedersen ’05, Carlin, Lobo & Viswanathan ’05, Schied &
• Solved by searching for Nash equilibrium.
Schöneborn ’08.
c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 65 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 66 / 74

Predatory trading and HF hot-potatos

One-period game model with A–Ch price impact

Assuming all Xi are deterministic this can be solved explicitly giving

n γ γ
− n+2 t t
Ẋi∗ (t) = αe η + βi e η ,

where
−n γ  − n γ T −1 x0

α= 1 − e n+2 η ,
n+2η n+1
−1  
γ  γη T x0
βi = e −1 Xi (T ) − Xi (0) +
η n+1

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 67 / 74
 Predatory trading and HF hot-potatos Predatory trading and HF hot-potatos

γ γ
Optimal strategies with n = 1, T = 1, η = 0.3 Optimal strategies with n = 1, T = 1, η = 20

100 100

80

50
60

40

0.2 0.4 0.6 0.8 1.0

20

0.2 0.4 0.6 0.8 1.0 !50

Distressed trader (blue) and one predator in a elastic market Distressed trader (blue) and one predator in an plastic market
(i.e. temporary impact > permanent impact) (i.e. permanent impact > temporary impact)

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 68 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 69 / 74

Predatory trading and HF hot-potatos

γ
Optimal strategies with n = 1, T = 1, η = 100

100

50

0.2 0.4 0.6 0.8 1.0

!50

Distressed trader (blue) and one predator in a highly plastic market.

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 70 / 74
 Predatory trading and HF hot-potatos Predatory trading and HF hot-potatos

Effect of predators, T = 1, x0 = 100, S0 = 100, Effect of predators, T = 1, x0 = 100, S0 = 100,

γ = η = 2% γ = 20 ∗ η = 2%
Comparison of n = 1 and n = 40 predators. Aggregated Holdings: Comparison of n = 1 and n = 40 predators. Aggregated Holdings:
100 100
100 100

80 80
50
50
60
60

40 0.2 0.4 0.6 0.8 1.0

40

20 0.2 0.4 0.6 0.8 1.0

!50
20

0.2 0.4 0.6 0.8 1.0

0.2 0.4 0.6 0.8 1.0 !20 !50 !100

Expected market price: Expected market price:

97.6 96.90
75
60.15
97.4
96.88

97.2

96.86 70 60.10
97.0

96.8 96.84
65 60.05
96.6
96.82

96.4

0.2 0.4 0.6 0.8 1.0 0.2 0.4 0.6 0.8 1.0
0.2 0.4 0.6 0.8 1.0 0.2 0.4 0.6 0.8 1.0

Expected execution cost E[C(X )]: 3.1% and 3.2% (compare with 3% when n = 0) Expected execution cost E[C(X )]: 33.3% and 40% (was 22% when n = 0)
Expected revenue per predator: 665 and 2.3.
Expected revenue per predator: 7.27 and 0.4.
Price and execution costs scale linearly with costs when keeping γη fixed.
c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 71 / 74 c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 72 / 74

Predatory trading and HF hot-potatos

HF hot-potato game

Schied & Zhang ’13 considered the following setup:

• two HF players X and Y trading in an Obizhaeva & Wang market
with G (t) = e−ρt
• trading at an equidistant discrete time grid
• with opposite initial positions X0 = −Y0 .
Using a Nash equilibrium analysis, they show that
• the optimal behaviour, if trading is frequent enough, involves a highly
oscillatory trading
• hot-potato effect with volume passed between traders
• the effect can be eliminated if transaction costs present and high
enough compared to LOB depth

c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 73 / 74
 Predatory trading and HF hot-potatos

Multi-agent setup summary

• Detailed analysis of market behaviour may require models with
interacting agents
• Mathematically, often done using game theory and searching for Nash
equilibria
• Predatory trading can be described as a game between one large seller
(prey) and n strategic traders (predators)
• Both from the theory and practice, we see that predators often first
trade in the same direction as the large trader leading to price
overshoot of which they then take advantage.
• The optimal behaviour highly dependent on the market characteristic
(e.g. which type of price impact dominates)
• More involved situations (e.g. strategic traders having longer trading
horizon) may lead to qualitatively different solutions
• Many other situations in which game analysis is interesting, e.g. high
trade volume (hot-potato) effect of trading between two agents.
c Jan Oblój, University of Oxford Price Impact Models and Market Microstructure 17 – 21 June 2019 74 / 74