Presented by: Thomas F. Coleman, Director, Waterloo Research Institute in Insurance, Securities and Quantitative Finance at University of Waterloo
MLX FinTech Conference II, Toronto, May 2018.
More info at: https://www.machinelearningx.net
MLX 2018 - Thomas F. Coleman, University of Waterloo
1. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Delta Hedging via Machine Learning
(w/o parametric models).*
Thomas F. Coleman
Director, WatRISQ
University of Waterloo
* Learning Minimum Variance Discrete Hedging Directly from the Market
Journal of Quantitative Finance, 2017. Ke Nian, Yuying Li, Thomas F. Coleman
2. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Seminar last week on Wall St : machine learning, text, arbitrage…
Speaker: Arun Verma, Bloomberg
“Machine Learning & Sentiment Analysis in Finance for Statistical
Arbitrage”
3. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Seminar last week on Wall St : machine learning, text, arbitrage…
4. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
5. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Delta Hedging via Machine Learning
(w/o parametric models).*
Thomas F. Coleman
Director, WatRISQ
University of Waterloo
* Learning Minimum Variance Discrete Hedging Directly from the Market
Journal of Quantitative Finance, 2017. Ke Nian, Yuying Li, Thomas F. Coleman
6. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
The Big Picture Message
Machine learning technology, w/o any finance knowledge or built-in
finance models, can (sometimes) outperform highly developed
specialized finance models in important (predictive) finance tasks.
7. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
The Big Picture Message
Machine learning technology, w/o any finance knowledge or built-in
finance models, can (sometimes) outperform highly developed
specialized finance models in important (predictive) finance tasks.
We illustrate this on one of the most important and well-studied problems
in finance : option hedging.
8. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
The Big Picture Message
Machine learning technology, w/o any finance knowledge or built-in
finance models, can (sometimes) outperform highly developed
specialized finance models in important (predictive) finance tasks.
We illustrate this on one of the most important and well-studied problems
in finance : option hedging.
The success we report is definitive.
9. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Hedging options is a major task in computational finance to remove
market risk .
Typical process (delta hedge):
1. Choose an option model
2. Calibrate the model (using real price(s)) to estimate model
parameters
3. Differentiate the model (with parameters now determined) at
current point to determine delta hedge
10. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Hedging options is a major task in computational finance to remove
market risk .
Typical process (delta hedge):
1. Choose an option model
2. Calibrate the model (using real price(s)) to estimate model
parameters
3. Differentiate the model (with parameters now determined) at
current point to determine delta hedge
BUT: - fails to capture parameter dependence on underlying
- does not minimize variance of hedge risk
11. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Practitioner's BS delta hedging for an option:
1. Assume that the underlying follows a BS underlying price model:
( )t t t tdS r q S dt S dW
12. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Practitioner's BS delta hedging for an option:
1. Assume that the underlying follows a BS underlying price model:
2. Calibrate BS implied volatility
from the market option price
( )t t t tdS r q S dt S dW
MKTV
( ; )BS MKTV S V
13. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Practitioner's BS delta hedging for an option:
1. Assume that the underlying follows a BS underlying model:
2. Calibrate BS implied volatility
from the market option price
3. Compute the BS delta using implied vol
( )t t t tdS r q S dt S dW
MKTV
( ; )BS MKTV S V
BSV
S
14. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
0In Coleman et al (2001) we illustrated the dependence of on S
15. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Minimum Variance Hedging
Goal: minimize option risk minimize variance (MV) in hedging error.
Let
: instantaneous change in underlying
: instantaneous change in option value
: position in underlying for a MV strategy
Hedging error:
MV delta hedging:
dS
dV
MV
MVdV dS
min variance ( ) (1)MV MVdV dS
16. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
For example, Hull and White [2017] directly attack (1) by assuming
Using market data to determine parameters a,b,c
Hull and White approximate
E( )BS BS
MV
V V
S S
2
E( ) ( )BS BS
a b c
S S T
17. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
For example, Hull and White [] directly attack (1) by assuming
Using market data to determine parameters a,b,c
Hull and White approximate
Improved results are obtained using this correction to BS-delta:
E( )BS BS
MV
V V
S S
2
E( ) ( )BS BS
a b c
S S T
2
( )
( )
BS BS
BS BS
MV
a b c
vega
S T
18. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Our new data-driven approach:
Solve
directly from the market data : no option models: use (standard) ML
techniques directly
min variance ( ) (1)MV MVdV dS
19. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Our new approach:
solve
directly from the market data : no option models: use (standard) ML
techniques directly
min variance ( ) (1)MV MVdV dS
learn
20. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
A Machine Learning Primer
Given m training points:
Where
x: feature vector
y: target (or outcome)
A regularized kernel method learns a nonlinear function
to capture the dependence between target y and feature vector x.
is then used to ‘predict’ y for feature selection x
1 1{( , ),...,( , )}m mx y x y
( )f x
( )f x
21. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
A Machine Learning Primer
Given m training points:
Where
x: feature vector
y: target (or outcome)
A regularized kernel method learns a nonlinear function
to capture the dependence between target y and feature vector x.
is then used to ‘predict’ y for feature selection x
Choose a positive definite kernel: e.g.,
1 1{( , ),...,( , )}m mx y x y
( )f x
( )f x
2
2
2
'
2
( , ')
x x
k x x e
22. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
A Machine Learning Primer
Given m training points:
Where
x: feature vector
y: target (or outcome)
A regularized kernel method learns a nonlinear function
to capture the dependence between target y and feature vector x.
is then used to ‘predict’ y for feature selection x
Choose a positive definite kernel: e.g.,
A regularized regression problem:
1 1{( , ),...,( , )}m mx y x y
( )f x
( )f x
2
2
2
'
2
( , ')
x x
k x x e
23. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
A regularized (nonlinear) regression problem (to determine (learn)
parameters that define ) :
2
1
1
min ( , ( )) ,
e.g., ( , ( )) ( )
m
f i i
i
p
i i i i
L y f x f
m
L y f x y f x
Defined by choice of K
f
24. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
A regularized (nonlinear) regression problem:
2
1
1
min ( , ( )) ,
e.g., ( , ( )) ( )
m
f i i
i
p
i i i i
L y f x f
m
L y f x y f x
25. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
A regularized (nonlinear) regression problem:
where m-vector minimizes the pos def quadratic:
2
1
1
min ( , ( )) ,
e.g., ( , ( )) ( )
m
f i i
i
p
i i i i
L y f x f
m
L y f x y f x
1
( ) ( , )
m
i i
i
f x k x x
2
1 1 1 1
min { ( ( , )) ( , )}m
m m m m
i j i j i j i j
i j i j
y k x x k x x
26. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
A regularized (nonlinear) regression problem:
where m-vector minimizes the pos def quadratic:
or
2
1
1
min ( , ( )) ,
e.g., ( , ( )) ( )
m
f i i
i
p
i i i i
L y f x f
m
L y f x y f x
1
( ) ( , )
m
i i
i
f x k x x
2
1 1 1 1
min { ( ( , )) ( , )}m
m m m m
i j i j i j i j
i j i j
y k x x k x x
2
2
min { }, where ( , )T
ij i jy K K K k x x
27. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
A regularized (nonlinear) regression problem:
where m-vector minimizes the pos def quadratic:
or
or
2
1
1
min ( , ( )) ,
e.g., ( , ( )) ( )
m
f i i
i
p
i i i i
L y f x f
m
L y f x y f x
1
( ) ( , )
m
i i
i
f x k x x
2
1 1 1 1
min { ( ( , )) ( , )}m
m m m m
i j i j i j i j
i j i j
y k x x k x x
2
2
min { }, where ( , )T
ij i jy K K K k x x
1
( )K I y
28. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Recipe for determining ML model
1. Choose feature vector x , kernel k
2. Get data
3. Form m-by-m matrix
4. Determine (cross validation)
5. Solve
( , ), 1:i ix y i m
( ( , ))i jK k x x
( )K I y
29. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Recipe for determining ML model
1. Choose feature vector x , kernel k
2. Get data
3. Form m-by-m matrix
4. Determine (cross validation)
5. Solve
Prediction:
( , ), 1:i ix y i m
( ( , ))i jK k x x
( )K I y
1
( ) ( , )
m
pred i i
i
y x k x x
30. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
To apply to our problem:
1. What is feature vector x?
2. Kernel function k?
3. Data?
4. Choose
And most importantly
31. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
To apply to our problem:
1. What is feature vector x?
2. Kernel function k?
3. Data?
4. Choose
And most importantly, what is objective function:
2
1
1
min ( , ( ))
m
f i i
i
L y f x f
m
min variance ( )MV MVdV dS
Key!
32. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Learn hedging function directly from the market using ML
approach: let be a positive definite kernel
where denotes change in (observed) market price
Solution is
( )f x
22
1
1
min ( ( ))
m
f i i i
i
V S f x f
m
t
i i iV V V
1
( ) ( , )
m
i i
i
f x k x x
( , ')k x x
33. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Learn hedging function directly from the market using ML
approach: let be a positive definite kernel.
where denotes change in (observed) market price
Solution is
Where solves the pos def. quadratic minimization:
Features x = moneyness , time-to-expiry;
( )f x
22
1
1
min ( ( ))
m
f i i i
i
V S f x f
m
t
i i iV V V
1
( ) ( , )
m
i i
i
f x k x x
( , ')k x x
2
1 1 1 1
min { ( ( ( , )) ( , )}m
m m m m
i i i i j i j i j
i i i j
V S k x x k x x
BS
( )
S
strike
34. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Empirical Results
Data: S&P 500 index option data from the OptionMetric Database
data: January 2, 2004 to August 31, 2015
35. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Empirical Results
Data: S&P 500 index option data from the OptionMetric Database
data: January 2, 2004 to August 31, 2015
Methods compared:
MV: Hull-White's minimum variance hedging based on formulation
LVF: MV hedging using correction based on LVF analysis
BS: implied volatility BS delta hedging
SABR : MV hedging from SABR stochastic volatility.
DKLSPL : direct learning a spline kernel hedging function
36. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Empirical Results
Data: S&P 500 index option data from the OptionMetric Database
data: January 2, 2004 to August 31, 2015
Methods compared:
MV: Hull-White's minimum variance hedging based on formulation
LVF: MV hedging using correction based on LVF analysis
BS: implied volatility BS delta hedging
SABR : MV hedging from SABR stochastic volatility.
DKLSPL : direct learning a spline kernel hedging function
Performance Measure:
2
1
2
1
( )
( )
m
M
i i
i
m
BS
i i
i
V S
GAIN
V S
37. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Empirical Testing
Data: S&P 500 index option data from the OptionMetric Database
data: January 2, 2004 to August 31, 2015
Sliding window:
3 years data 1 mth-out-of sample
learning model testing
9 buckets, 9 models
38. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
9 buckets, 9 models
39. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
40. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
41. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
42. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Concluding remarks
We have proposed to learn the hedge function directly from market data
(using ML techniques).
43. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Concluding remarks
We have proposed to learn the hedge function directly from market
data (using ML techniques).
Empirical results show
- proposed approach definitively outperforms existing MV hedge
methods in daily hedging
44. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Concluding remarks
We have proposed to learn the hedge function directly from market data
(using ML techniques).
Empirical results show
- proposed approach definitively outperforms existing MV hedge
methods in daily hedging
- proposed approach definitively outperforms implied vol BS hedge
in daily, weekly and monthly hedging
45. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
Concluding remarks
We have proposed to learn the hedge function directly from market data
(using ML techniques).
Empirical results show
- proposed approach definitively outperforms existing MV hedge
methods in daily hedging
- proposed approach definitively outperforms implied vol BS hedge
in daily, weekly and monthly hedging
Question: What other important (predictive) problems in finance,
traditionally solved with highly developed specialized finance models,
are amenable to general data driven machine learning techniques?
46. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018
REFERENCES
T. F. Coleman, Y. Kim, Y. Li, A. Verma, Dynamic hedging with a deterministic
local volatility function model, The Journal of Risk 5 (2001) 63:89.
J. Hull, A. White, Optimal delta hedging for options Journal of Banking &
Finance, 82:180-190, 2017
G. Bakshi, C. Cao, Z. Chen, Empirical performance of alternative option pricing
models, The Journal of Finance 52 (1997) 2003.
S. Crepey, Delta-hedging vega risk, Quantitative Finance 4 (2004) 559{579.
G. Bakshi, C. Cao, Z. Chen, The Journal of Finance 52 (1997) 2003{2049.
K. P. R., R. Schenk-Hoppe, C.-O. Ewald, Risk minimization in stochastic
volatility models: model risk and empirical performance, Quantitative Finance 9
(2009) 693{704.
G. Bakshi, C. Cao, Z. Chen, The Journal of Finance 52 (1997) 2003.
J. Hutchinson, A. Lo, T. Poggio. A nonparametric approach to pricing
and hedging derivative securities via learning networks. The Journal
of Finance. 49(3):851-889, 1994
47. Thomas Coleman , University of Waterloo Machine Learning Fintech Conference 2.0. May 2, 2018