Download Commodities Futures Price Prediction An Artificial Intelligence

Commodities Futures Price
Prediction An Artificial
Intelligence Approach
Thesis Defense
Commodities Markets
• Commodity
– A good that can be processed and resold
– Examples – corn, rice, silver, coal
• Spot Market
• Futures Market
Futures Markets
• Origin
• Motivation
– Hedgers
• Producers
• Consumers
– Speculators
• Size and scope
– CBOT (2002)
• 260 million contracts
• 47 different products
Profit in the Futures Market
• Information
– Supply
•
•
•
•
Optimal production
Weather
Labor
Pest damage
– Demand
• Industrial
• Consumer
• Time series analysis
Time Series Analysis Background
• Time Series – examples
–
–
–
–
–
–
River flow and water levels
Electricity demand
Stock prices
Exchange rates
Commodities prices
Commodities futures prices
• Patterns
Time Series Analysis - Methods
•
•
•
•
•
Linear regression
Non-linear regressions
Rule based systems
Artificial Neural Networks
Genetic Algorithms
Data
•
•
•
•
•
Daily price data for soybean futures
Chicago Board of Trade
Jan. 1, 1980 – Jan. 1, 1990
Datastream
Normalized
Why use an Artificial Neural
Network (ANN)?
• Excellent pattern recognition
• Other uses of ANN and financial time series
analysis
– Estimate generalized option pricing formula
– Standard & Poors 500 index futures day trading
system
– Standard & Poors 500 futures options prices
ANN Implementation
• Stuttgart Neural Network Simulator, version 4.2
• Resilient propagation (RPROP)
– Improvement over standard back propagation
– Uses only the sign of the error derivative
• Weight decay
• Parameters
–
–
–
–
Number of inputs 10 and 100
Number of hidden nodes 5, 10, 100
Weight decay 5, 10, 20
Initial weight range +/- 1.0, 0.5, 0.25, 0.125, 0.0625
ANN Data Sets
• Training set Jan. 1, 1980 – May 2, 1983
• Testing set May 3, 1983 – Aug. 29, 1986
• Validation set Sept. 2, 1986 – Jan. 1, 1990
ANN Results
• Mean Error
– 100 input
• 12.00
• 24.93
– 10 input
• 10.62
• 25.88
– Cents per bushel
Why Evolve the parameters of an
ANN?
• Selecting preferred parameters is a difficult
poorly understood task
• Search space is different for each task
• Trial and error is time consuming
• Evolutionary techniques provide powerful
search capabilities for finding acceptable
network parameters.
Genetic Algorithm Implementation
• Galib, version 4.5 (MIT)
• Custom code to implement RPROP with weight
decay
• Real number representation
–
–
–
–
–
–
Number of input nodes (1 – 100)
Number of hidden nodes (1 – 100)
Initial weight range (0.0625 – 2.0)
Initial step size (0.0625 – 1.0)
Maximum step size (10 – 75)
Weight decay (0 – 20)
Genetic Algorithm –
Implementation (continued)
•
•
•
•
Roulette wheel selection
Single point crossover
Gausian random mutation
High mutation rate
Evaluation Function
• Decode the parameters and instantiate a
network using them
• Train the ANN for 1000 epochs
• Report the lowest total sum of squared error
for both training and testing data sets
• Fitness equals the inverse of the total error
reported.
Parameter Evolution - Results
• GANN Mean error 10.82
• NN Mean error 10.62
• Conclusions:
– GANN performance is close and out performs the
majority of networks generated via trial and error
– Genotype / Phenotype issue
– Other, possibly better GA techniques
• Multipoint crossover
• Tournament selection
Evolving the Weights of an ANN
• Avoid local minima
• Avoid tedious trial and error search for
learning parameters
• Perform search of broad, poorly understood
solution space and maximize the values for
function parameters
Weight evolution Implementation
•
•
•
•
•
•
Galib, version 4.5 (MIT)
Custom written neural network code
Real number representation
Gausian Mutation
Two point crossover
Roulette wheel selection
Weight Evolution – objective
function
• Instantiate a neural network with the weight
vector (I.e. the individual)
• Feed one epoch of the training data
• Fitness equals the inverse of the sum of the
squared network error returned
Weight Evolution – keeping the
best individual
• Fitness function evaluates against training
set only
• Objective function evaluates against
training set as well, but only for retention of
candidate best network
• Meta-fitness, or meta-elite individual
Weight Evolution - Results
• Mean Error
– GANN-Weight 10.67
– GANN 10.82
– NN 10.61
• Much faster
• Fewer man hours
Summary
• Pure ANN approach is very man hour
intensive and expert experience is valuable
• Evolving network parameters requires few
man hours, but many hours of
computational resources.
• Evolving the network weights provides
most of the performance for smaller cost in
both human and computer time