prediction of the foreign exchange market using classifying neural network doug moll chad zeman
TRANSCRIPT
Prediction of the Foreign Exchange Market Using Classifying Neural Network
Doug Moll
Chad Zeman
The Problem
Using neural networks, can we predict future
foreign exchange rates to profit from short-
term fluctuations?
Outline
Project History Project Proposal Data Set MPL Results PNN Results
Project History
Senior Seminar
Used Trajan for regression networks
Attempted to predict direction & movement size
Less than desirable results
Proposal
Classification of Up or Down movement
Continue to use Trajan
Maintain same biases to compare to previous
research
Minimize time to learn new tool
Data Set
Inputs (1994 – 2003)
Percent change of CA/US exchange rate
Interest differential of short term interest rates
(CA – US)
Lagged Values
Exchange Rate data 94-03
94-03 CA/US rate
1.25
1.3
1.35
1.4
1.45
1.5
1.55
1.6
1.65
1994 1995 1996 1997 1998 1999 2000 2001 2002 2003
Ex
ch
an
ge
Ra
te
CA/US Exchange Rate
Percent Change Model
CA/US Percent Weekly Change
-4%
0%
4%
1994 1995 1996 1997 1998 1999 2000 2001 2002 2003
% C
han
ge
CA/US %Change
Data Set
Equalize training cases
Same number of Up examples as Down for
each of the three data periods
Training
Verification
Testing
Trajan Algorithm
User Defined Settings
Inputs & Outputs
Training, Verification, Test data splits
Network Type
Trajan Algorithm
Automatically Determined Settings
Network Complexity (# of hidden nodes)
Trajan Algorithm
Randomly builds networks
Trains using backpropagation
Utilize cross-verification techniques
Evaluate networks based on verification error
Cross-reference with out-of-sample test data
Results – MLP - Daily
16 inputs
20 hidden nodes
1 output
0.3 momentum
0.1 learning rate 50 epochs
Results – MLP - Daily
Data Set Performance
Training 58.41%
Verification 53.59%
Testing 53.59%
Results – MLP - Weekly
16 inputs
22 hidden nodes
1 output
0.3 momentum
0.1 learning rate 4 epochs
Results – MLP - Weekly
Data Set Performance
Training 52.73%
Verification 62.50%
Testing 55.47%
Probabilistic Neural Networks
Finite deterministic network Three layers
PNN Example – Training Example A
Exchange Rate
Interest Rate
1.35
2.5%
Input Layer
A
Target Output = Up
PatternLayer
OutputLayer
Up
Down
Pattern Layer
Receives input vector Calibrates Gaussian bell
PNN Example – Training Example A
Exchange Rate
Interest Rate
1.35
2.5%
Input Layer
A 100%
Target Output = Up
PatternLayer
OutputLayer
Up
Down
A
1.352.5%
1
PNN Example – Training Example B
Exchange Rate
Interest Rate
1.40
2.7%
Input Layer
A
100%
Target Output = Down
PatternLayer
OutputLayer
Up
Down
A
1.352.5%
B
1.402.7%
1
PNN Example – Out-of-Sample
Exchange Rate
Interest Rate
1.39
2.7%
Input Layer
A
80%
PatternLayer
OutputLayer
Up
Down
A
1.352.5%
B
1.402.7%
.40
.10 20%
Results – PNN - Daily
16 inputs
1224 hidden nodes
2 outputs
Results – PNN - Daily
Data Set Performance
Training 54.82%
Verification 51.14%
Testing 51.96%
Results – PNN - Weekly
16 inputs
256 hidden nodes
2 outputs
Results – PNN - Weekly
Data Set Performance
Training 77.73%
Verification 54.69%
Testing 57.03%
Conclusions
Predicting foreign exchange market is a tough problem
PNN vs. MLP
Weekly vs. Daily data