Forecasting direction of EUR/USD monthly exchange rate using support vector machine UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus G53IDS 09/10 Presenter: Li Junlong Supervisor: Ho Sooi Hock
28
Embed
Forecasting EUR/USD monthly exchange rate using support vector machine
This is my 2009 presentation on development of a forecast machine.
Welcome message from author
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
Forecasting direction of EUR/USD monthly exchange rate using support vector machine
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
G53IDS 09/10
Presenter: Li Junlong
Supervisor: Ho Sooi Hock
2
Project Objective
• Apply SVM to forecast the nominal monthly exchange rate of EUR/USD
• Review SVM forecast performance
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
3
What does the software do?
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
Learning Forecasting Analysis
4
How is this useful?
• For government– Defense against possible impending financial disaster
• Imagine foreseeing the next currency crisis
• For private– Trading decisions based on good market predictions
tend to perform better• Make money
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
5
What is forecasting?
• It is to estimate, predict or calculate in advance, through the use of various forecasting methods, to determine the future expectations.
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
6
Which kind of forecast methodology?
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
SVM
7
Why FOREX, not stocks?
• Less insider information compared with stocks.• Seems to be more closely related to public
information.– such as ∆ of GDP growth, balance of payment, interest
rate, inflation rate, etc.
• Need more complete package information to make better forecasts.– The information package of FOREX is more complete as
many are publicly available.
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
8
Why EUR/USD?
• According to the Bank for International Settlements study in 2007, the most heavily traded products on the spot market were:
• EUR/USD was chosen because of its popularity.
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
EUR/USD27%
USD/JPY13%
GBP/USD12%
Others48%
9
Why monthly forecast?
• Many statistics are released on monthly basis, such as ∆ of US & Europe inflation rate, balance of payment etc. – Interpolation function for certain data released longer than
monthly basis• such as quarterly release of %∆ of GDP.
– Aggregation function needed for certain data released shorter than monthly basis
• such as ∆ of price.– Combination of both function for certain data released without
fixed timings• such as ∆ of central bank discount rate.
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
10
Why not shorter periods?
• The FX market is relatively nosier in very short term.– Greater exposure to errors
• Economic policies usually require certain period of time for results to be observable, due to the problem of time lags.– Recognition lag– Decision lag– Effect lag
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
11
Why direction, not value?
• Limitation of SVM, a binary classifier.• Higher complexity to predict value.• Less probability of error
– Possibilities of direction = 2– Possibilities of value ≈ ∞
• “Trading driven by a certain forecast with a small forecast error may not be as profitable as trading guided by an accurate prediction of the direction of movement.” (W. Huang et al. 2005)
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
12
Past related researches?
• Stocks– Autoregressive Integrated Moving Average (ARIMA)– Random Forests / Decision Trees (DT)– Artificial Neural Networks (ANN)– Support Vector Machine (SVM)– Support Vector Regression (SVR)
• FX– Autoregressive Integrated Moving Average (ARIMA)– Artificial Neural Networks (ANN)– Support Vector Regression (SVR)
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
13
Why SVM, not others?
• DT learners tend to create over-complex trees that do not generalize the data well. This is called overfitting.[1]
• NN learners does not produce unique results.[2]• Empirical tests suggest that SVMs tend to
produce lower error rate in stock forecasting compared to ANN and ARIMA methods.[3]
• SVR is used for variable forecasts.
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
14
What is the expected accuracy?
• Wei Huang, Yoshiteru Nakamori, Shou-Yang Wang, “Forecasting stock market movement direction with support vector machine”, October 2005– Hit ratio of SVM ≈ 73%– Hit ratio of Combining model ≈ 75%
• Jingtao Yao, Chew Lim Tan, “A Case Study on Using Neural Networks to Perform Technical Forecasting of Forex”, 2000.– Hit ratio of ANN ≈ 70%
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
15
Where are the evidences?
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
16
What is SVM?
• Method of supervised learning for machines• Learn to linearly classify data
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
17
How 'bout non-linearly separable data?
• Feature map the input space to a usually high dimensional feature space where the data points become linearly separable, called kernels– Polynomial– Radial Basis Function
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
18
How SVM works?
Note: Above are just simplified explanations, actual work is much more complicated.
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
19
Demo?
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
20
Plans for implementation?
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
21
What are your input data?
• Any data that has been generally observed to show strong relationship with the direction of EUR/USD exchange rate– Backed by evidence. (rare)*****– Backed by research publications. (handful)***– Theories based on challengeable assumptions which
may or may not hold true in the real world. (many)*
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
22
Where do you get your data?
• Internet– Government websites
• U.S. Bureau of Economic Analysis• U.S. Bureau of Labor Statistics• Eurostat of European Commission• The European Central Bank
– Private organizations• Exchange-Rates.org• Google Finance
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
• Financial theories based on generalization, observable correlations may not be perfect.
• Financial theories are hard to verify because the movement of market is constantly being influenced by other factors.
• Choosing the market determinants can be seen as a qualitative art.
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
24
What are the challenges?
• Knowledge– High complexity of the theories.– Large amount of theories from different fields.– Steep learning curve for me.
• Experience– First time
• Doing something you have never done before is always harder than doing something that you already have some experience of.
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
25
What are the drawbacks of SVM?
• Does not assess probabilistic confidence of classification.
• Only provides binary classification.
• Counterarguments• Multiclass SVM though combinations of multiple SVMs• SVM for regression was proposed in 1996 by Vladimir
Vapnik, Harris Drucker, Chris Burges, Linda Kaufman and Alex Smola.[4]
• More accurate to predict direction than certain value.[5]
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
26
“
Humans are essentially complex bio-machines.
With the constant increasing complexity of man-
made machines, someday they will be able to learn
and make decisions independently without the need
of their creators.
”
Lesson learnt?
UNIMKL-004363, School of Computer Science, Faculty of Science, The University of Nottingham Malaysia Campus
27
• [1]Bramer, Max. Principles of Data Mining. London : Springer, 2007. ISBN 978-1-84628-765-7.• [2],[3],[5] Wei Huang, Yoshiteru Nakamori, Shou-Yang Wang, “Forecasting stock market movement direction with support
vector machine”, Computers & Operations Research, Volume 32, Issue 10, Applications of Neural Networks, October 2005
• [4] Harris Drucker, Chris J.C. Burges, Linda Kaufman, Alex Smola and Vladimir Vapnik (1997). "Support Vector Regression Machines". Advances in Neural Information Processing Systems 9, NIPS 1996, 155-161, MIT Press.
• KIM, Kyoung-jae, 2003. Financial time series forecasting using support vector machines, Neurocomputing, Volume 55, Issues 1-2 (September 2003), Pages 307-319.
• CAO, L. J. and Francis E. H. TAY, 2003. Support Vector Machine With Adaptive Parameters in Financial Time Series Forecasting, IEEE Transactions on Neural Networks, Volume 14, Issue 6, November 2003, Pages 1506-1518.
• CAO, Lijuan and Francis E. H. TAY, 2001. Financial Forecasting Using Support Vector Machines. Neural Computing & Applications, Volume 10, Number 2 (May 2001), Pages 184-192.
• YANG, Haiqin, Laiwan CHAN and Irwin KING, 2002. Support Vector Machine Regression for Volatile Stock Market Prediction. In: Intelligent Data Engineering and Automated Learning: IDEAL 2002, edited by Hujun Yin, et al., pages 391--396, Springer.
• CAO, L. J. and Francis E. H. TAY, 2000. Feature Selection for Support Vector Machines in Financial Time Series Forecasting. In: Intelligent Data Engineering and Automated Learning - IDEAL 2000: Data Mining, Financial Engineering, and Intelligent Agents , edited by Kwong Sak Leung, Lai-Wan Chan and Helen Meng, pages 268-273.
• CHEN, Wun-Hua, Jen-Ying SHIH and Soushan WU, 2006. Comparison of support-vector machines and back propagation neural networks in forecasting the six major Asian stock markets, International Journal of Electronic Finance, Volume, Issue 1, pages 49-67.