Empirical Game-Theoretic Analysis of Algorithmic Trading ...
Transcript of Empirical Game-Theoretic Analysis of Algorithmic Trading ...
![Page 1: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/1.jpg)
Empirical Game-Theoretic Analysis of Algorithmic Trading Scenarios
Michael WellmanLynn A. Conway Collegiate Professor of Computer Science & Engineering
University of Michigan
![Page 2: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/2.jpg)
Modeling Algorithmic Trading• Research agenda• Implications on market performance• Separate distinct algo roles & tactics• Evaluating market and regulation designs
Example studies:• Latency arbitrage and
frequent call markets• Strategic market choice• Market making• Strategic shading• Market manipulation
This talk:• Spoofing• Trend following and
market stability
• Approach: Agent-based simulation─ Comprehensive modeling of financial
market ecosystem─ Capturing heterogeneity, microstructure,
fine-grained information flows─ Combine with game-theoretic analysis
![Page 3: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/3.jpg)
Levels of Agent (ARB-BOT) Behavior1. Passive search for arbitrage opportunities2. Attempts to amplify arb opps through
purposeful instigation of market movements (e.g., spoofing)
3. Attempts to create new arb opps• new financial instruments• deliberate fragmenting
4. Malicious subversion of markets
most aggressive
![Page 4: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/4.jpg)
Level 2: Market Manipulation
• SEC Definition: Intentional conduct designed to deceive investors by controlling or artificially affecting the market for a security• Spoofing• Dodd-Frank defn: bidding or offering with the intent to cancel the bid or offer
before execution• Widely known strategies, several recent prosecutions
• Recent work: Agent-Based Model of Spoofing• Wang & W., AAMAS-17• Key idea: Rational investors learn from market prices under normal
conditions, vulnerable to spoofing
![Page 5: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/5.jpg)
Spoofing refers to the practice of submitting large spurious orders to buy or sell some security to mislead other traders’ beliefs.
![Page 6: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/6.jpg)
Source: Financial Conduct Authority, Animated Example of Mr Coscia’s Trading
True buy order
True sell order
Spoofing sell order
Spoofing buy order
![Page 7: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/7.jpg)
Spoofing Study
Motivating Questions
• Can we model a market where spoofing effectively manipulates rational traders’ beliefs about prices?
• What is the impact of spoofing on market performance and trading behavior? Can we quantify the effect of spoofing?
Xintong Wang
![Page 8: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/8.jpg)
6,80,25,1…
Estimating an Empirical Game
![Page 9: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/9.jpg)
Agent Strategic Choices - EGTA (HBL, ZI)
0.0
0.2
0.4
0.6
0.8
1.0
HBL
Ado
ptio
n Ra
te
No spoofing
A1 A2 A3 B1 B2 B3 C1 C2 C30.0
0.2
0.4
0.6
0.8
1.0
HBL
Ado
ptio
n R
ate
No spoofing
A1 A2 A3 B1 B2 B3 C1 C2 C3
When spoofing is absent, investors generally have incentives to make bidding decisions based on order book information.
(a) N = 28 (b) N = 65
![Page 10: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/10.jpg)
HBL Improves Price Discovery - EGTA (HBL, ZI) vs EGTA (ZI)
120
130
140
150
160
170
180
190
200
RM
SD
HBL and ZI ZI
A1 A2 A3 B1 B2 B3 C1 C2 C3 90
100
110
120
130
140
150
160
170
RMSD
HBL and ZI ZI
A1 A2 A3 B1 B2 B3 C1 C2 C3
(a) N = 28 (b) N = 65
![Page 11: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/11.jpg)
16500
16700
16900
17100
17300
17500
17700
17900
18100
18300
18500
Surp
lus
HBL and ZI ZI
A1 A2 A3 B1 B2 B3 C1 C2 C339500
40000
40500
41000
41500
42000
42500
43000
43500
44000
Surp
lus
HBL and ZI ZI
A1 A2 A3 B1 B2 B3 C1 C2 C3
(a) N = 28 (b) N = 65
HBL Improves Market Surplus - EGTA (HBL, ZI) vs EGTA (ZI)
![Page 12: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/12.jpg)
A Spoofable Market with HBL Traders
(a) N = 28 (b) N = 65
Start SpoofingStart Spoofing
Amplified (diminished) spoofing effect in markets with more HBL (ZI).
![Page 13: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/13.jpg)
Surplus Redistribution with Spoofing
-300
-250
-200
-150
-100
-50
0
50
100
A1 A1 A2 A3 B1 B2 B2 B3 C3
Surp
lus
HBL surplus difference ZI surplus difference
-600
-500
-400
-300
-200
-100
0
100
A1 A1 A1 A2 A2 A3 B1 B2 B2 B3 C1 C2 C2 C3
Surp
lus
HBL surplus difference ZI surplus difference
(a) N = 28 (b) N = 65
In a market with spoofing, ZI benefits from HBL’s spoofed beliefs.
![Page 14: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/14.jpg)
Agent Strategic Choices with Spoofing
(a) N = 28 (b) N = 65
When re-equilibrating games with spoofing, HBL often remains in mixed equilibria but with smaller proportions.
0.0
0.2
0.4
0.6
0.8
1.0
HBL
Ado
ptio
n R
ate
No spoofing Spoofing
A1 A2 A3 B1 B2 B3 C1 C2 C3 0.0
0.2
0.4
0.6
0.8
1.0
HBL
Ado
ptio
n Ra
te
No spoofing Spoofing
A1 A2 A3 B1 B2 B3 C1 C2 C3
![Page 15: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/15.jpg)
Market Surplus with Spoofing
(a) N = 28 (b) N = 65
The presence of spoofing generally decreases market surplus.
16500
17000
17500
18000
18500
Surp
lus
No spoofing with HBL Spoofing with HBLNo spoofing without HBL Spoofing without HBL
A1 A2 A3 B1 B2 B3 C1 C2 C3 39500
40500
41500
42500
43500
Surplus
No spoofing with HBL Spoofing with HBL Spoofing without HBL
A1 A2 A3 B1 B2 B3 C1 C2 C3
![Page 16: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/16.jpg)
Spoofing: Ongoing Research Questions
1. What are some of the general characteristics of market environments subject to spoofing?
2. Are there more robust ways for exchanges to disclose order book information?• Yes: cloaking mechanisms [Wang, Vorobeychik, W., IJCAI-18]
3. Are there strategies by which traders can exploit order book information but in less vulnerable ways?• Yes: price blocking or offset strategies [Wang, Hoang, W., ICML-
19 workshop]
4. Based on a model of this kind, can we design effective spoofing detection methods?
![Page 17: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/17.jpg)
EGTA Investigation of Financial Market Stability• Under what conditions might a market shock be amplified?• Trend-following agents• A simple “technical” trading strategy• Can be profitable given market frictions• Effect on market stability and performance
Erik Brinkman
![Page 18: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/18.jpg)
Effect of Price Shock in a Model where Agents Observe Current Fundamental
Price - μShock: exogenous insertion of large sell order
Market stabilizes quickly, as agent beliefs unaffected by prices
![Page 19: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/19.jpg)
Effect of Shock with Noisy Observation of Fundamental
Price - μ
Agent beliefs influenced by prices, so shock has more persistent effect
![Page 20: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/20.jpg)
Market Frictions Contribute to Mispricing
Fundamental Mispricing
Suggests long trends signal profit opportunities
![Page 21: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/21.jpg)
High Frequency Trend Follower
1. Continually looks for monotonic trends in transaction prices (fundamental shift)
2. Take best price in direction of trend, and resubmit for a more extreme price
3. Withdraw if un-transacted for a long period of time
TF strategy present in equilibrium, given model with noisy observations
![Page 22: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/22.jpg)
Shock with Trend Followers
Price - μ
![Page 23: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/23.jpg)
TFs Reduce Mispricing during Trends
Fundamental Mispricing
Equilibrium comparisons
![Page 24: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/24.jpg)
TFs Crowd Out Use of Transaction Price Information
Fraction using transaction information
![Page 25: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/25.jpg)
Crowding Out Changes Shock Response
Price - μ
![Page 26: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/26.jpg)
Trend Following Exacerbates Shock Impact
Shock Magnitude
Max Difference
![Page 27: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/27.jpg)
Trend Following Helps Recovery
RMSD
Shock Magnitude
![Page 28: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/28.jpg)
TFs Reduce Market Efficiency
Without TFs 50.7%Without TFs 50.3%With TFs 40.6%
Why?• Crowding out use of price information• Increasing adverse selection
![Page 29: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/29.jpg)
Conclusions: Trend Followers
& Market Stability
• TF can be rational given standard market frictions• Imperfect fundamental information• Delayed market access
• TF affect on market performance• Less price-awareness• More adverse selection• Reduced efficiency
• TF affect on shock response • Amplified magnitude, shorter duration• Representative of flash crashes
![Page 30: Empirical Game-Theoretic Analysis of Algorithmic Trading ...](https://reader031.fdocuments.us/reader031/viewer/2022020915/619dcfa36a93fb79460480bb/html5/thumbnails/30.jpg)
EGTA & Algorithmic Trading: Conclusions
Closing Thoughts• Algorithmic trading at the leading
edge for application of autonomous agents• Is algorithmic manipulation the next
frontier?• Is stability at risk?• Can AI be part of the solution?
• Computational strategic reasoning a useful tool for understanding financial settings, design for market quality and stability
Our other AI/Finance Study Areas• (more on algorithmic trading)
• Financial credit networks• Strategic payment routing• Intermediation• Network formation
• Bank leverage regulation