Optimization of Trading Rules

Overview
Optimization of Trading Rules

with a Penalty Term for Increased Risk-Adjusted
Risk
Risk-Adjusted
Performance
Trading rules
Performance Evaluation
Benchmarks
Constrained Optimization
Empirical results
Thomas Hellstr
m
Hellstr
Hellstrm
Department of Computing Science
Ume
Ume University
Sweden
email: thomash@
@cs.
thomash
cs.umu.se
umu.se
thomash@cs.umu.se
www.cs
cs..umu.se/~
thomash
www.
umu.se/~thomash
www.cs.umu.se/~thomash
Thomas Hellstrm 2001
Trading Rules
Example of a Trading Rule
are used for decision support:
g can often be parameterized as g[X], and optimized w.r.t X
Level of Resistance
Example of a Trading Rule
if the price brakes through the resistance: g(t)=1 Buy
50 and 100 day moving averages
180
180
SELL
160
120
100
80
BUY
SELL
60
120
100
80
BUY
60
BUY
40
40
BUY
20
0
140
g(t)
SELL
SELL
Stock price y(t)
140
g(t)
Stock price y(t)
160
200
400
1
600
t
800
1000
20
-1
0
5
200
400
600
t
800
1000
-1
6
Trading Channel Breakout

180
mavx1+x2
160
BUY
140
120
moving
average
mavx1
BUY
100
80
g(t)
Stock price y(t)
Gaussian Volume
if y(t) penetrates mavx1+x2 : g(t)=1 Buy
BUY
60
40
1
20
0
200
400
600
t
800
1000
-1
7
Combining Moving Averages with Volume
Trading Channel Breakout
The crossing moving averages is combined with the Gaussian volume into
one compound trading rule mav:
Level of Resistance
10
Performance Evaluation
Performance measures are needed twice:
1) In the learning phase when optimal parameters
are determined
2) When the trading rules are tested out of sample
Two common ways to compute performance:
Profit in Simulated Trading
Hit rate at a fixed prediction horizon
In this paper the Hit rate is used.
11
12
Hit Rate
Benchmarks
The positive hit rate for a Buy rule is the fraction of

buy signals, which are followed by an increase in stock price:
A benchmark should provide an alternative and

standardized way to produce predictions.
The Naive Prediction of Return asserts todays return R h (t )

s
as the best estimate of R h ( t + h ) . For a time period [1,,T]
and a set of stocks S, the h-day positive hit rate for the
naive return predictor is computed as
The negative hit rate for a Sell rule is the fraction of

sell signals, which are followed by a decrease in stock price
13
Benchmarks
14
Optimizing the Trading Rules

The function g is normally parameterized with a few
parameters X that can be determined to optimize performance on
the training data.
The notation g[X] denotes this parameterization.
The Naive prediction asserts todays price Closes(t)

as the best estimate of Closes(t+h). For a time period [1,,T]
and a set of stocks S, the h-day positive hit rate for the
Naive- predictor is computed as
The trading rule normally issues Buy and Sell signals only for a
minor part of the time steps. This is bad for two reasons:
1) Bad statistical significance for the performance
2) Risk for over optimization. I.e: bad generalization
15
16
Reformulation to a Smooth Problem
A Constrained Optimization Problem

We therefore formulate a constrained optimization problem for a
Buy rule g :
XL and XH are lower and upper bounds for the unknown

parameters X and the other constraint is the total number
of generated Buy signals. Using a hard constraint leads to a
non-smooth problem. Furthermore, it is hard to decide on a
crisp value for N0 .
The constraint acts as a regularizer since the search space for

the function J is reduced by requiring a minimum number of
trading signals.
17
18
Sliding Windows
Optimization
Since we can
cant use cross validation
90
91
92
93
94
95
96
97
year
The optimization problem is a box-bounded nonconvex global optimization problem, where no

derivates are available.
optimization evaluation
Run 1
Run 2
Run 2
Run 4
Run 5
Run 6
In this paper we are using the DIRECT algorithm by

Jones (1993). The algorithm estimates the Lipschitz
constant and uses it to control the trade-off between
global versus local search.
The performance is computed as the average

performance for the 6 runs
19
Results
20
Example of Optimized Trading Rules for 1992
The 32 largest Swedish stocks have been used in the tests:
Regularized
trading
rules
Non
regularized
Non
regularized
Each year gets different optimal rules
The non-regularized optimization over-fits data and is no better

than the bench marks out-of-sample. The regularized rules are all
significantly better than the bench marks.
21
References
Stability of the Found Optima
Optimization of Trading Rules with a Penalty Term for Increased

Risk-Adjusted Performance.
Published in Advanced Modeling and Optimization, 2(3):135-149, 2000.
Global Optimization of Costly Nonconvex Functions, with Financial
Applications. (with Kenneth Holmstrm. Theory of Stochastic Processes
2000.
Parameter Tuning in Trading Algorithms Using ASTA.
(with Kenneth Holmstrm). Published in Y.S. Abu-Mostafa, B.LeBaron,
A.W. Lo, and A.S. Weigend (Editors), Computational Finance 1999,
p. 343-357.
ASTA - a Tool for Development of Stock Prediction Algorithms.
Published in Theory of Stochastic Processes, 5(21)(1-2):22-31, 1999.
ASTA - User's Reference Guide.
Technical Report UMINF-00.16 ISSN-0348-0542, Department of
Computing Science Ume University, Ume Sweden, 2000.
D.R. Jones, C.D. Perttunen, and B.E. Stuckman. Lipschitzian
optimization without the Lipschitz constant. Journal of Optimization Theory
and Applications, 79(1):157--181, October 1993.
The stability and the relevance of the found optima is also tested.
The trading rules for 1992 are applied not only for 1992 but also
for the following years up to 1997:
The average performance is lower than for the year-by-year

optimized rules (Hte).
22
23
email: thomash@
thomash@cs.
cs.umu.se
umu.se
www.cs
thomash
www.cs..umu.se/~
umu.se/~thomash
24

Optimization of Trading Rules

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Optimization of Trading Rules

Uploaded by

Copyright:

Available Formats

Overview

Optimization of Trading Rules

Thomas Hellstrm 2001

Thomas Hellstrm 2001

Example of a Trading Rule

are used for decision support:

g can often be parameterized as g[X], and optimized w.r.t X

Thomas Hellstrm 2001

Thomas Hellstrm 2001

Example of a Trading Rule

if the price brakes through the resistance: g(t)=1 Buy

50 and 100 day moving averages

Stock price y(t)

Stock price y(t)

Thomas Hellstrm 2001

Trading Channel Breakout

Stock price y(t)

if y(t) penetrates mavx1+x2 : g(t)=1 Buy

Thomas Hellstrm 2001

Combining Moving Averages with Volume

Trading Channel Breakout

Thomas Hellstrm 2001

Thomas Hellstrm 2001

Thomas Hellstrm 2001

Thomas Hellstrm 2001

The positive hit rate for a Buy rule is the fraction of

A benchmark should provide an alternative and

The Naive Prediction of Return asserts todays return R h (t )

The negative hit rate for a Sell rule is the fraction of

Thomas Hellstrm 2001

Thomas Hellstrm 2001

Optimizing the Trading Rules

The Naive prediction asserts todays price Closes(t)

Thomas Hellstrm 2001

Thomas Hellstrm 2001

Reformulation to a Smooth Problem

A Constrained Optimization Problem

XL and XH are lower and upper bounds for the unknown

The constraint acts as a regularizer since the search space for

Thomas Hellstrm 2001

The optimization problem is a box-bounded nonconvex global optimization problem, where no

In this paper we are using the DIRECT algorithm by

The performance is computed as the average

Thomas Hellstrm 2001

Example of Optimized Trading Rules for 1992

The 32 largest Swedish stocks have been used in the tests:

Each year gets different optimal rules

The non-regularized optimization over-fits data and is no better

Thomas Hellstrm 2001

Stability of the Found Optima

Optimization of Trading Rules with a Penalty Term for Increased

The average performance is lower than for the year-by-year

Thomas Hellstrm 2001

Thomas Hellstrm 2001

You might also like