Non-fixed and Asymmetrical Margin Approach to Stock Market Prediction using Support Vector Regression Haiqin Yang, Irwin King and Laiwan C han Department of Computer Science and Engineering The Chinese University of Hong Kong November 18-22, 2002 ICONIP’02
25
Embed
Haiqin Yang, Irwin King and Laiwan Chan Department of Computer Science and Engineering
Non-fixed and Asymmetrical Margin Approach to Stock Market Prediction using Support Vector Regression. Haiqin Yang, Irwin King and Laiwan Chan Department of Computer Science and Engineering The Chinese University of Hong Kong November 18-22, 2002 ICONIP ’ 02. Index. Motivation. - PowerPoint PPT Presentation
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Non-fixed and Asymmetrical Margin Approach to Stock Market
Prediction using Support Vector Regression
Haiqin Yang, Irwin King and Laiwan ChanDepartment of Computer Science and Engineering
The Chinese University of Hong Kong
November 18-22, 2002ICONIP’02
2/25
Index
Motivation
SVR Introduction Approach
Conclusion
Experiments & Results
3/25
Motivation
Combine them:
Non-fixed and Asymmetrical margin
Two characteristics: fixed and symmetrical
Predictive accuracy only?
Downside risk!
4/25
Support Vector Regreesion (SVR) introduction Developed by Vapnik (1995)Developed by Vapnik (1995)
The objective function f is represented by the dotted points.
6/25
Related Applications
Support Vector Method for Function Approximation, Support Vector Method for Function Approximation, Regression Estimation and Signal Processing (VapniRegression Estimation and Signal Processing (Vapnik et al., 1996)k et al., 1996)
Predicting time series with support vector machine Predicting time series with support vector machine (Muller et al., 1997)(Muller et al., 1997)
Application of support vector machines to financial tiApplication of support vector machines to financial time series forecasting (E.H.Tay and L.J.Cao. 2001)me series forecasting (E.H.Tay and L.J.Cao. 2001)
7/25
Approach
Two characteristics: 4 kinds of margins
SSymmetricalymmetrical AsymmetricalAsymmetrical
FFixedixed
Non-fixedNon-fixed
fixed,
symmetrical.
FASMFASM
NASMNASM
FAAMFAAM
NAAMNAAM
+ + + + + + + + + +
+ + + + + + + + + +
+ + + + + + + + + +
+ + + + + + + + + +
8/25
Previous setting
Previous others’ method
SSymmetricalymmetrical AsymmetricalAsymmetrical
FFixedixed
Non-fixedNon-fixed
SSymmetricalymmetrical AsymmetricalAsymmetrical
FFixedixed FASMFASM FAAMFAAM
Non-fixedNon-fixed NASMNASM NAAMNAAM
In our previous work: Support Vector Machines Regression for volatile stock market prediction (IDEAL’02)
9/25
New Approach
Two characteristics of the margin in
– insensitive loss function: fixed and symmetrical.
Non-fixedNon-fixed
AsymmetricalAsymmetricalSSymmetricalymmetrical
FFixedixed
10/25
Formulas A general type of –Insensitive loss function
Fixed and Symmetrical Margin (FASM):
Fixed and Asymmetrical Margin (FAAM):
Non-fixed and Symmetrical Margin (NASM):
Non-fixed and Asymmetrical Margin (NAAM):
up margin
down margin
11/25
Formulas
QP problem:
s.t. Objective function:
Kernel function:
e.g. RBF
12/25
How to set margin?
Margin width:Up margin:Down margin:
13/25
Experiment
Accuracy Metrics• MAE:
• UMAE:
• DMAE:
• actual value,
• predictive value
• number of testing data
m
iii pa
m 1
||1
m
paiii
ii
pam ,1
)(1
m
paiii
ii
apm ,1
)(1
iaipm
Total error
Upside risk
Downside risk
14/25
Experiment Description
Model: Data: Hang Seng Index (HSI),
Dow Jones Industrial Average (DJIA). Time periods: Jan. 2, 1998 ~ Dec. 29, 2000 (3 years) Ratio of training data and testing data: 5:1. Procedures: one day ahead prediction. Environments
• CPU: Pentium 4, 1.4 G
• Memory: RAM 512M
• OS: Windows2000
• Time: few hours.
15/25
Experiment Description
Three kinds of experiments• Test the effect of parameters in NAAM to obtain a
better result.
• Compare the result of NAAM with NASM, AR(4), RBF network (also test the effect of the number of hidden units).
• Compare the results of NAAM, NASM with FASM and FAAM.