order flows and execution costs - university of california...
Post on 18-Feb-2020
4 Views
Preview:
TRANSCRIPT
1
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Order Flows and Execution Costs
Mike Ludkovski
UC Santa Barbara
IPAM Workshop on High Frequency Trading, April 14, 2015joint w/Kyle Bechler, Sebastian Jaimungal
Ludkovski Execution & Order Flow
1
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Optimal Execution Modeling
Typical OE models are “reduced-form”Ingredients include
I reference price (e.g. Brownian motion)I realized price (e.g. reference + offset) ;I price impact specification (e.g. transient linear impact w/exponential resilience);I liquidity state, etc.,
Trading is continuous: diffusion settingScheduling time-scale: 5–10 minutes
Ludkovski Execution & Order Flow
2
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Data
Actual data is based on the Limit Order BookFull information about all ticker events and snapshots of the bookMathematically represented in terms of point processes or queuesTime-scale of� 1 second
How to reconcile/connect these frameworks?
Ludkovski Execution & Order Flow
3
Order Flow Data Analysis Optimal Execution Numerical Illustrations
OE vs LOB
Existing OE models are oblivious to LOB statesDespite these valuable proposals to model the dynamics of the order book at small time scales, no direct use of such
effects in an optimal trading framework is currently available. —CA Lehalle, Handbook of Systemic Risk 2013
Typical OE algorithms are “open-loop” – calibrate and forget
Wish to come up with a higher-level picture of LOB at the time-scale of the schedulerto build a “closed-loop” system:LOB→ Liquidity/Market State→ OE algorithm
Ludkovski Execution & Order Flow
3
Order Flow Data Analysis Optimal Execution Numerical Illustrations
OE vs LOB
Existing OE models are oblivious to LOB statesDespite these valuable proposals to model the dynamics of the order book at small time scales, no direct use of such
effects in an optimal trading framework is currently available. —CA Lehalle, Handbook of Systemic Risk 2013
Typical OE algorithms are “open-loop” – calibrate and forgetWish to come up with a higher-level picture of LOB at the time-scale of the schedulerto build a “closed-loop” system:LOB→ Liquidity/Market State→ OE algorithm
Ludkovski Execution & Order Flow
4
Order Flow Data Analysis Optimal Execution Numerical Illustrations
What is Liquidity?
Challenges:How to measure liquidity of the LOB?How to measure price impact via the LOB?How to reconcile Market Orders and Limit Orders in LOB?
Static snapshots of the LOB have been extensively studiedFor example short-term effects from book imbalance and “micro price” which in turndrives mid-price movementsLiquidity can be statically measured in terms of
I Bid-Ask SpreadI DepthI Execution Cost
But these features are too detailed/fleeting at the OE scaleDynamic view is provided by the order flows (time-scale of minutes)
Ludkovski Execution & Order Flow
4
Order Flow Data Analysis Optimal Execution Numerical Illustrations
What is Liquidity?
Challenges:How to measure liquidity of the LOB?How to measure price impact via the LOB?How to reconcile Market Orders and Limit Orders in LOB?
Static snapshots of the LOB have been extensively studiedFor example short-term effects from book imbalance and “micro price” which in turndrives mid-price movementsLiquidity can be statically measured in terms of
I Bid-Ask SpreadI DepthI Execution Cost
But these features are too detailed/fleeting at the OE scaleDynamic view is provided by the order flows (time-scale of minutes)
Ludkovski Execution & Order Flow
5
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Outline
Aim: include LOB info into OE schedulersPART I: Initial empirical analysis with a view to explain what is (statistically) importantfor order executionPART II: A simple model to incorporate order flow into OEFor another take, see Incorporating Order-Flow into Optimal Execution by Cartea andJaimungal (on SSRN, 2015)
Ludkovski Execution & Order Flow
6
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Order Flow
Need to aggregate LOB data to understand the more persistent features relevant forOEMoving from microstructure to the meso-scale (complements line of research thataggregates from point processes to diffusions)Aggregation naturally leads to consideration of Order FlowsRelated to metrics such as Market Toxicity or Liquidity StateThere is a lot of evidence that flows are persistent (hours or even days)
⇒ Trading shows up as a disturbance of order flow – links to price impact
Ludkovski Execution & Order Flow
7
Order Flow Data Analysis Optimal Execution Numerical Illustrations
LOB Evolution
●
0 50 100 150 200
010
0020
0030
0040
0050
0060
00
1
TV.1
4.t[s
pan[
1], 9
]
●
●
●●
● ●
●
●
● ● ●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● 51.6
051
.65
51.7
051
.75
Tick-by-tick evolution of LOB queues at the touch. Market orders are in orange. Mid price isplotted at the top (TEVA, May 3, 2011)
Ludkovski Execution & Order Flow
8
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Main Aspects of Order Flow
Must separately treat Limit Orders and Market OrdersConsequently, must also separately treat each side of the book (cross-effects)Aggregating by time-slicing or volume-slicing?Order Flow has intrinsic links to price behavior: mechanical (slippage) andinformational (price discovery)Economically, OF is tied to informational costs:
I Participants fear that other traders are better informedI Noise traders trade in all directions; informed traders are deliberate⇒ When order flow is toxic market makers provide less liquidityI Example of a metric: VPIN by Easley, Lopez de Prado, O’Hara (2012abcd) – ties toxicity
to flow imbalance = ratio of noise/informed traders
Ludkovski Execution & Order Flow
9
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Empirical Order Flow−
4e
+0
5−
3e
+0
5−
2e
+0
5−
1e
+0
50
e+
00
MSFT 06/21/2012
Cu
mu
lativ
e E
xecu
ted
Ord
ers
Figure: Cumulative net volumeof MSFT executed trades overJune 21, 2012
10 11 12 13 14 15 16
0.0
00
.01
0.0
20
.03
0.0
4
Time
Pri
ce
Im
pa
ct
80
,00
0 S
ha
res
Figure: Price Impact for MSFTover same day
0 50 100 150 200 250 300
−0
.2−
0.1
0.0
0.1
0.2
Volume Buckets
Ne
t O
rde
r F
low
Im
ba
lan
ce
Figure: MSFT Net Order Flowover buckets of 1,000,000
Ludkovski Execution & Order Flow
10
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Notation
LOB consists of pairs (pjt , v
jt ) indexed from the top queue.
Mid-price: PM(t) := (pA1 (t) + pB
1 (t))/2Market orders (MO) and Limit orders (LO) form asynchronous time-series:time-stamp Ti , signed volume Oi , limit price Si
To synchronize & aggregate, construct bucketsSlice the ticker according to intervals defined by (τk ) (e.g. equal volume)Market buys: VMA
k =∑
i:τk≤T Mi ≤τk+1
OMi 1{OM
i >0}.
MIk = VMAk − VMB
k
Contemporaneous ask limit orders at the touch:
VLAk =
∑i:τk≤T L
i ≤τk+1
OLi 1{SL
i =pB1 (T
Li )}
Ludkovski Execution & Order Flow
11
Order Flow Data Analysis Optimal Execution Numerical Illustrations
How to Model Price Impact
Permanent Price ImpactTransient Price Impact
1 Large orders walk the book (depends on the LOB shape)2 Liquidity providers replenish the queues and the book bounces back (resiliency)
Obizhaeva/Wang: Realized execution price: Pt = Pt + Dt
Unaffected P is arithmetic BMResilience D decays exponentially to zeroEach term is linearly impacted by trading:
Pt = P0 + λMIt + σWt , Dt = e−ρtD0 + νMIt .
Assuming Gaussian D0, gives a conditional Gaussian distribution for ∆P, linear in MIt
Ludkovski Execution & Order Flow
11
Order Flow Data Analysis Optimal Execution Numerical Illustrations
How to Model Price Impact
Permanent Price ImpactTransient Price Impact
1 Large orders walk the book (depends on the LOB shape)2 Liquidity providers replenish the queues and the book bounces back (resiliency)
Obizhaeva/Wang: Realized execution price: Pt = Pt + Dt
Unaffected P is arithmetic BMResilience D decays exponentially to zeroEach term is linearly impacted by trading:
Pt = P0 + λMIt + σWt , Dt = e−ρtD0 + νMIt .
Assuming Gaussian D0, gives a conditional Gaussian distribution for ∆P, linear in MIt
Ludkovski Execution & Order Flow
12
Order Flow Data Analysis Optimal Execution Numerical Illustrations
LOB Perspective
Immediate effect of a trade is deterministic given the LOB state. Cannot distinguishvarious sub-effectsDifficult to measure long-term effectsHave to describe the dynamic impact on LOB – eg arrival/cancellation rates
Large Tick AssetsConcentrate on liquid US equitiesTick size is $0.01; asset price is $20-$50Spread is almost always 1 tickThe two queues at the touch are deep (> 10AES)Orders rarely “walk the book” and there is no shape to the LOBTypical immediate price impact is zeroexamples: MSFT, INTC, TEVA
Ludkovski Execution & Order Flow
12
Order Flow Data Analysis Optimal Execution Numerical Illustrations
LOB Perspective
Immediate effect of a trade is deterministic given the LOB state. Cannot distinguishvarious sub-effectsDifficult to measure long-term effectsHave to describe the dynamic impact on LOB – eg arrival/cancellation rates
Large Tick AssetsConcentrate on liquid US equitiesTick size is $0.01; asset price is $20-$50Spread is almost always 1 tickThe two queues at the touch are deep (> 10AES)Orders rarely “walk the book” and there is no shape to the LOBTypical immediate price impact is zeroexamples: MSFT, INTC, TEVA
Ludkovski Execution & Order Flow
13
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Price Impact of Market Flow
−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5
−0.10
−0.05
0.00
0.05
0.10
Market Order Imbalance
∆ P ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
● ●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
● ●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
● ●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
● ●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
x 105
Figure: Relationship between mid-price change ∆P and net market order flow MI acrossvolume-aggregated slices of 150,000 shares (1/80 of ADV) for MSFT during Jan-Mar 2011. Theaverage price change was computed using Loess regression.
Ludkovski Execution & Order Flow
14
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Explaining Price Movement based on Flows
One sided flow consumes liquidity asymmetrically so the price movesLots of buying→ price goes up(Extreme MI typically anticipates low ∆P – observed S-shape)BUT: Even under balanced flow the price can make significant movesGood state: MI = 50,000 and ∆P = 0.Bad state: MI = 20,000 and ∆P = 0.04 large price move based on small net volume.Potential causes of such weak LOB resilience: are:
1 Illiquid LOB2 Cancellations of limit orders (vanishing liquidity)
Ludkovski Execution & Order Flow
15
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Static Liquidity of the LOB
Spread: pA1 − pB
1 . Always 1-2 ticks – not indicativeStatic volume imbalance: |vA
1 (t)− vB1 (t)| - jumps each
time the mid-price movesDepth:
∑i v j
i (e.g. top 2 levels of the book)Price Impact: total effect relative to touch ifinstantaneously sell M shares:
PIi := M−1iM∑
i=1
vi (Si−p1)+viM (SiM−p1), iM = min{i :∑j<i
vj > M}
Vo
lum
e
𝑆𝑝𝑟(𝑡)
𝑝4𝐴
Price
Type equation here.
𝑝3𝐴 𝑝2
𝐴 𝑝1𝐴
𝑃𝑀 𝑝1
𝐵 𝑝2𝐵
𝑣1𝐴
𝑝3𝐵 𝑝4
𝐵
𝑣2𝐴
𝑣4𝐴
𝑣3𝐴
𝑣1𝐵
𝑣2𝐵
𝑣3𝐵
𝑣4𝐵
Ludkovski Execution & Order Flow
16
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Vanishing Liquidity
Static depth can also be misleadingOne-sided market flows can induce wild swingsin additions/cancellations and generate a lot ofmid-price volatilityThis depends on the correlation between VL amdMO and further latent regimesPicture from Lehalle et al (2012) – July 19, 2012
Ludkovski Execution & Order Flow
17
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Limit Order Flows
Limit order flow should generally be positive (moreadditions than cancellations) – counteracts marketordersInstances where VL is negative are highly indicativeof major price movesThis is not surprising from the LOB mechanicsBut shows that time series of limit flows is crucial toidentify reduced liquidity/toxicity
Net Order Flow at the Touch
Fre
qu
en
cy
−2e+05 0e+00 2e+05 4e+05
02
00
40
06
00
80
01
00
0
Ludkovski Execution & Order Flow
18
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Limit Order Flows
−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5
−3−2
−10
12
3
Market Order Imbalance
VLB
x 105
x 105
−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5
−3−2
−10
12
3
Market Order Imbalance
VLA
x 105
x 105
Figure: Relationship between net market order flow MI and concurrent top-level limit flow VLj forMSFT during Jan-Mar 2011. Left: Bid, right: Ask. Highlighted points indicate buckets where theprice change exceeded 5 ticks.
Ludkovski Execution & Order Flow
19
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Can We Predict Scarce Liquidity?
Generalized additive model to explain ∆P(GAM) Logistic regression to predict large price moves aka scarce liquidityMost important predictor is MO flowIn both cases LO have much more explanatory power than price impact or bookimbalance.This is not actionable (LO is contemporaneous)There is also some memory of scarce liquidity (based on ACF for volume-sliced timeseries)
Ludkovski Execution & Order Flow
20
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Dependence Between Limit & Market Orders
●
0 50 100 150 200
01
00
02
00
03
00
04
00
05
00
06
00
0
1
TV.1
4.t
[sp
an
[1],
9]
●
●
●●
● ●
●
●
● ● ●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
51
.60
51
.65
51
.70
51
.75
−ve Corr (TEVA)
0 50 100 150 200 250 300 350
050
0010
000
1500
0
Ticker EventsP
oste
d Vo
lum
e
●●●
●●●
●●●●●
●●●●●●
●●●●
●●●●
●●●●
●
●●●
●
●
●●●●●●●●●●●●●●●●●●●●●●●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●
●●●●
●
●
●
●●
●
●●
●
●
●●●●
●●
●
●●●●●●●●●●●●●●●●●●●●●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●●
●
●
●
●
●●●
●
●
●
●
●●
●●
●
●
●●●●
●
●●●
●
●
●
●
●
●
● ●●
●
●
●
●●
●
●●●● ●●●
●●●●●
●●●●●
●
●
●●
●
●
●
●
●
●●
●
●●●
●
●●
●
●●●
●
●
●●
● ●●
●
●●
●●●●●●
●●●●●●●
●●
●●
●●●
●●●●●
●●
●●
●●
●●
●●
●● ●
●●
●●●●●●●
●●●
●●●●●●●●●●●●
●●●●●●
●●
●
●●●●●●
●●●●
●●●●
●●●●
●●●●●●●●●
●●
●●●
●
●●●
●●●●●●
●
36.8
2036
.830
36.8
4036
.850
+ve Corr (ORCL)
Ludkovski Execution & Order Flow
21
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Dependence Between Limit & Market Orders
Could represent via regimes: resilient (VL and MI positive correlation)Scarce: negative correlation (price fade/"front running")Limit order flow is a key factor to explain price move conditional on MOOpen Problem: How to define model-free dependence of 4 asynchronous markedpoint processes?
Ludkovski Execution & Order Flow
22
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Correlation Between Limit & Market Orders
12 13 14 15 16
−0
.20
.00
.20
.40
.60
.8
Hours from Midnight
Co
rre
latio
n
11 12 13 14 15 16
−0
.4−
0.2
0.0
0.2
0.4
Time
Co
rre
latio
n
Smoothed contemporaneous correlation between VL and MI over the past 2.5 hours using30-second buckets (4 days in 2011). Left: TEVA. Right: MSFT. Solid: bid-side. Dashed: ask-side
Ludkovski Execution & Order Flow
23
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Another View of Persistency in Liquidity
−20000 −10000 0 10000 20000
−0.2
−0.1
0.00.1
0.2
Market Order Imbalance
Price C
hange
Figure: ∆P against MI and accompanying linear best fit lines for Feb 10, 2011 and March 9, 2011in TEVA. Higher correlation implies stronger MI impact (also see Cartea & Jaimungal, 2015).
Ludkovski Execution & Order Flow
24
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Current Work
Quantify observed stylized features using statistical models (nonparametricregression)Suggest some ways of measuring the temporal correlation between limit and marketorder flowsInvestigate existence of statistical regimes (hidden Markov factors??)
Ludkovski Execution & Order Flow
25
Order Flow Data Analysis Optimal Execution Numerical Illustrations Two-Step Approximation
Part II: Incorporating OF into OE
Two main concerns when executing a large order:1 Price Impact
Spatial effect: consume liquidity in terms of standing limit orders
2 Information LeakageTemporal effect: executed trade appears on the ticker tapeOther traders react and adjust their behavior (eg “front-running”) – will receive worseprice in the future
AIM: integrate these ideas into a dynamic optimal execution frameworkTreat order flow as a (controlled) state variable
Ludkovski Execution & Order Flow
25
Order Flow Data Analysis Optimal Execution Numerical Illustrations Two-Step Approximation
Part II: Incorporating OF into OE
Two main concerns when executing a large order:1 Price Impact
Spatial effect: consume liquidity in terms of standing limit orders2 Information Leakage
Temporal effect: executed trade appears on the ticker tapeOther traders react and adjust their behavior (eg “front-running”) – will receive worseprice in the future
AIM: integrate these ideas into a dynamic optimal execution frameworkTreat order flow as a (controlled) state variable
Ludkovski Execution & Order Flow
26
Order Flow Data Analysis Optimal Execution Numerical Illustrations Two-Step Approximation
Literature Landscape
Work with continuous trading ratesAssume a parametric form for inventory risk (Gatheral-Schied)Risk neutral agents (Gueant & Lehalle, ...)No fill risk (in contrast to Cartea & Jaimungal, Gueant et al, Bayraktar-L, ... )Treat only market orders (in the future: hybrid models like in Guilbaud & Pham,Carmona & Webster, ... )Postulate one-sided trading (in the future: allow two-sided to consider potential formarket manipulation)Flat (constant depth) LOB→ linear price impactFlat LOB→ quadratic instantaneous execution cost(Order book resilience (Obizhaeva & Wang, Alfonsi et al))(Empirical evidence: Farmer, Bouchaud, ... )
Ludkovski Execution & Order Flow
27
Order Flow Data Analysis Optimal Execution Numerical Illustrations Two-Step Approximation
Contributions
Introduce order flow imbalance as a state variableSuggest a simple mechanism for informational costs (inspired by ELO12) – temporalimpact beyond usual spatial impactWill also endogenize the execution horizon T – can trade faster under favorableconditionsDerive closed-form approximate strategies for the resulting optimization problem
Ludkovski Execution & Order Flow
28
Order Flow Data Analysis Optimal Execution Numerical Illustrations Two-Step Approximation
Execution Model
Inventory xt : liquidate x0
t 7→ xt is absolutely continuous; trading rate is αt
dxt = −αtdtUnaffected price St : martingale
Order flow imbalance Yt :I Mean-reverting to zeroI Stationary in long-runI Affected by execution algorithms
Yt > 0: buyers-market; Yt < 0: sellers market
Ludkovski Execution & Order Flow
28
Order Flow Data Analysis Optimal Execution Numerical Illustrations Two-Step Approximation
Execution Model
Inventory xt : liquidate x0
t 7→ xt is absolutely continuous; trading rate is αt
dxt = −αtdtUnaffected price St : martingaleOrder flow imbalance Yt :
I Mean-reverting to zeroI Stationary in long-runI Affected by execution algorithms
Yt > 0: buyers-market; Yt < 0: sellers market
Ludkovski Execution & Order Flow
29
Order Flow Data Analysis Optimal Execution Numerical Illustrations Two-Step Approximation
Order Flow Dynamics
Unaffected order flow: dY 0t = −βY 0
t dt + σ dWt
With execution:dYt = (−βYt − φ(αt )) dt + σ dWt
i.e. Yt = Y 0t +
∫ t0 e−β(t−s)αsds
Typical cases:I φ(α) = φt (deterministic information cost)I φ(α) = ηα (linear in trading rate)
Assume that flow is independent of price (empirical relationship is not clear) - moreon this later
Ludkovski Execution & Order Flow
30
Order Flow Data Analysis Optimal Execution Numerical Illustrations Two-Step Approximation
Optimization Problem
Objective: v(x , y) := infα∈A Ex ,y
[∫ T00 g(αs) + λ(xs) + κY 2
s ds]
Realized horizon T0 := inf{t : xαt = 0} – endogenous to the strategy α
g(α): price impact
g(α) = α2 (constant-depth LOB)
λ(x): inventory risk
λ(x) = cx2 (Almgren-Chriss criterion) / λ(x) = cx (similar to Gatheral-Schied)
κY 2 : information cost
Unbalanced order flow: Higher liquidity costs
Ludkovski Execution & Order Flow
30
Order Flow Data Analysis Optimal Execution Numerical Illustrations Two-Step Approximation
Optimization Problem
Objective: v(x , y) := infα∈A Ex ,y
[∫ T00 g(αs) + λ(xs) + κY 2
s ds]
Realized horizon T0 := inf{t : xαt = 0} – endogenous to the strategy αg(α): price impactg(α) = α2 (constant-depth LOB)
λ(x): inventory risk
λ(x) = cx2 (Almgren-Chriss criterion) / λ(x) = cx (similar to Gatheral-Schied)
κY 2 : information cost
Unbalanced order flow: Higher liquidity costs
Ludkovski Execution & Order Flow
30
Order Flow Data Analysis Optimal Execution Numerical Illustrations Two-Step Approximation
Optimization Problem
Objective: v(x , y) := infα∈A Ex ,y
[∫ T00 g(αs) + λ(xs) + κY 2
s ds]
Realized horizon T0 := inf{t : xαt = 0} – endogenous to the strategy αg(α): price impactg(α) = α2 (constant-depth LOB)λ(x): inventory riskλ(x) = cx2 (Almgren-Chriss criterion) / λ(x) = cx (similar to Gatheral-Schied)
κY 2 : information cost
Unbalanced order flow: Higher liquidity costs
Ludkovski Execution & Order Flow
30
Order Flow Data Analysis Optimal Execution Numerical Illustrations Two-Step Approximation
Optimization Problem
Objective: v(x , y) := infα∈A Ex ,y
[∫ T00 g(αs) + λ(xs) + κY 2
s ds]
Realized horizon T0 := inf{t : xαt = 0} – endogenous to the strategy αg(α): price impactg(α) = α2 (constant-depth LOB)λ(x): inventory riskλ(x) = cx2 (Almgren-Chriss criterion) / λ(x) = cx (similar to Gatheral-Schied)κY 2 : information costUnbalanced order flow: Higher liquidity costs
Ludkovski Execution & Order Flow
31
Order Flow Data Analysis Optimal Execution Numerical Illustrations Two-Step Approximation
HJB Equation
0 = −βYvY + 12σ
2vYY + infα≥0{g(α)− αvx − φ(α)vY}+ κY 2 + λ(x)
Finite-fuel boundary condition: v(0, y) = 0 for all yNonlinear parabolic PDEHard to understand the structurePositivity constraint on α is challenging
To gain insights: build approximating problems by(i) solving the fixed-horizon problem(ii) optimizing over T
Ludkovski Execution & Order Flow
31
Order Flow Data Analysis Optimal Execution Numerical Illustrations Two-Step Approximation
HJB Equation
0 = −βYvY + 12σ
2vYY + infα≥0{g(α)− αvx − φ(α)vY}+ κY 2 + λ(x)
Finite-fuel boundary condition: v(0, y) = 0 for all yNonlinear parabolic PDEHard to understand the structurePositivity constraint on α is challengingTo gain insights: build approximating problems by
(i) solving the fixed-horizon problem(ii) optimizing over T
Ludkovski Execution & Order Flow
32
Order Flow Data Analysis Optimal Execution Numerical Illustrations Two-Step Approximation
Fixed Horizon Problemu(T , x , y) = inf
(αt )∈A(T ,x)Ex ,y
[∫ T0 α2
s + λ(xαs ) + κY 2s ds
]HJB equation becomes
uT =12σ2uyy + κy2 + λ(x)− βyuy + inf
α
{α2 − αux − ηαuy
}(1)
Singular boundary condition: limT↓0 u(T , x , y) =∞ if x 6= 0
Proposition
The solution of (1) has the form
u(T , x , y) = x2A(T ) + y2B(T ) + xyC(T ) + D(T ), (2)
where A,B,C,D solve a matrix Riccati ordinary differential equation.
Note: Riccati equations parameterized in terms of time-to-maturity τ
Ludkovski Execution & Order Flow
33
Order Flow Data Analysis Optimal Execution Numerical Illustrations Two-Step Approximation
Execution Speed
The corresponding optimal rate of liquidation is
αDt =
xt (2A(τ) + ηC(τ)) + Yt (C(τ) + 2ηB(τ))
2.
Execution rate is linear in xt and in Yt (generalizes Almgren-Chriss)The Proposition only treats the unconstrained case α ∈ R: if T is large relative to x0or Yt is negative enough then αD < 0As t → T , the dynamic trading rate stabilizes, resembling a VWAP strategy.
Ludkovski Execution & Order Flow
34
Order Flow Data Analysis Optimal Execution Numerical Illustrations Two-Step Approximation
Myopic Strategies (φ(α) = φt )
Suppose agent myopically optimizes only against price impact:uM(T , x , y) := inf(xt )
(∫ T0 x2
s + λ(xs)ds)
+∫ T
0 κEy [Y 2s ]ds =: I +O
If λ(x) = cx2 solution is
xMH
t =x sinh(
√c(T − t))
sinh(√
cT )
αMHt =
√cx cosh(
√c(T − t))
sinh(√
cT )
Now φt = ηαMH
t is the above deterministic function of t → Yt is Gaussian with knownmomentsIMH(T , x) =
√cx2 coth(
√cT )
OMH(T , x , y) = κ∫ T
0
(ye−βt −
∫ t0 e−β(t−s)ηαMH
s ds)2
+ σ2
2β (1− e−2βt ) dt
Similarly have closed-form expressions for cases λ(x) = cx (Quadratic) and λ(x) = 0(VWAP)Next step: Take existing closed-form expressions uM and optimize over T
Ludkovski Execution & Order Flow
34
Order Flow Data Analysis Optimal Execution Numerical Illustrations Two-Step Approximation
Myopic Strategies (φ(α) = φt )
Suppose agent myopically optimizes only against price impact:uM(T , x , y) := inf(xt )
(∫ T0 x2
s + λ(xs)ds)
+∫ T
0 κEy [Y 2s ]ds =: I +O
If λ(x) = cx2 solution is
xMH
t =x sinh(
√c(T − t))
sinh(√
cT )
αMHt =
√cx cosh(
√c(T − t))
sinh(√
cT )
Now φt = ηαMH
t is the above deterministic function of t → Yt is Gaussian with knownmomentsIMH(T , x) =
√cx2 coth(
√cT )
OMH(T , x , y) = κ∫ T
0
(ye−βt −
∫ t0 e−β(t−s)ηαMH
s ds)2
+ σ2
2β (1− e−2βt ) dt
Similarly have closed-form expressions for cases λ(x) = cx (Quadratic) and λ(x) = 0(VWAP)Next step: Take existing closed-form expressions uM and optimize over T
Ludkovski Execution & Order Flow
35
Order Flow Data Analysis Optimal Execution Numerical Illustrations Two-Step Approximation
Optimizing the Horizon
T ∗ = arg minT u(T , x , y)
Lemma: T ∗ ∈ (0,∞) (closed-form for T 7→ u(T , x , y))Open-loop (static): find T ∗ at the outset and implement αt (T ∗(x , y), xt ,Yt )
Closed-loop (dynamic): continuously recompute T ∗:
αMt (x , y) := αM(T ∗(xt ,Yt ), xt ,Yt )
Realized horizon T0(x , y) becomes randomDynamically recomputing T ∗ - adapt to changing Yt without the indefinite horizonfinite-fuel problem
Next up: we show these are in fact good approximations!
Ludkovski Execution & Order Flow
35
Order Flow Data Analysis Optimal Execution Numerical Illustrations Two-Step Approximation
Optimizing the Horizon
T ∗ = arg minT u(T , x , y)
Lemma: T ∗ ∈ (0,∞) (closed-form for T 7→ u(T , x , y))Open-loop (static): find T ∗ at the outset and implement αt (T ∗(x , y), xt ,Yt )
Closed-loop (dynamic): continuously recompute T ∗:
αMt (x , y) := αM(T ∗(xt ,Yt ), xt ,Yt )
Realized horizon T0(x , y) becomes randomDynamically recomputing T ∗ - adapt to changing Yt without the indefinite horizonfinite-fuel problem
Next up: we show these are in fact good approximations!
Ludkovski Execution & Order Flow
36
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Comparison of costs
Optimal Execution Costv uD uML uD uML
E[J(α)] 4.257 4.264 4.317 4.483 4.547SD(J(α)) 1.50 1.45 1.39 1.77 1.84E[T0] 3.87 3.70 3.48 3.43 3.43
Legend:
v : directly from the HJB pde (fully numerical)
u(x , y): closed-loop optimization of T ∗; T0 is random
u(T ∗, x , y): static optimization T0 = T ∗
uML: VWAP on [0,T ∗]
D superscript refers to dynamic strategies based on Proposition 1 (Riccati eqns)
Ludkovski Execution & Order Flow
37
Order Flow Data Analysis Optimal Execution Numerical Illustrations
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5
Trad
ing
Rat
e α
t
αt*
α~tDL
α~tML
α(2)ML
αtDL
αML
0.6
0.8
1.0
1.2
1.4
1.6
1.8
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5
0.0
0.5
1.0
1.5
2.0
2.5
3.0
Inve
ntor
y
αtx
α~tD
α~tMyoLin
αtD
αMyoLin
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5
Time t
Orde
r Flow
Imba
lance
(Yt0 )
−0.4
−0.2
0.0
Figure: Comparison of trading rates (αt ) for each of the six strategies along the shown simulated path of (Y 0t ) (The
realized (Yt ) depends on the strategy chosen).
Ludkovski Execution & Order Flow
38
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Execution Paths
Inve
ntor
y x t
0 1 2 3 4 5
0.0
0.5
1.0
1.5
2.0
2.5
3.0
0 1 2 3 4 5
Time t
Ord
er F
low
Imba
lanc
e (Y
t)
−0.6
−0.4
−0.2
0.0
0.2
0.4
Figure: Top: 200 simulated trajectories from strategy αDt . Highlighted are three trajectories resulting from different
Yt -paths. Bottom: Corresponding realizations of t 7→ Yt .
Ludkovski Execution & Order Flow
39
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Extensions
How to estimate/filter/infer Yt from LOB data?information leakage φ(α) depends on Yt?Correlated St and Yt : dependence between price movement and order flow
Thank You!
Ludkovski Execution & Order Flow
39
Order Flow Data Analysis Optimal Execution Numerical Illustrations
Extensions
How to estimate/filter/infer Yt from LOB data?information leakage φ(α) depends on Yt?Correlated St and Yt : dependence between price movement and order flow
Thank You!
Ludkovski Execution & Order Flow
40
Order Flow Data Analysis Optimal Execution Numerical Illustrations
References
A. Cartea, S. JaimungalIncorporating Order-Flow into Optimal Executionhttp://ssrn.com/abstract=2557457 (2015)
J. Gatheral and A. SchiedDynamical models of market impact and algorithms for order executionhttp://papers.ssrn.com/sol3/papers.cfm?abstract_id=2034178 (2013).
C.-A. LehalleMarket Microstructure knowledge needed to control an intra-day trading processHandbook on Systemic Risk, J-P Fouque and J Langsam Eds (2013)
K. Bechler and M. LudkovskiOptimal Execution with Dynamic Order Flow ImbalancePreprint, (2014) http://arxiv.org/abs/1409.2618
Ludkovski Execution & Order Flow
top related