Evolving Market Index Trading Rules

1
EvoNet Summer School 2001
Evolving Market Index Trading Rules
Atif Azad Laura Dipietro Steven Gustafson Robin
Purshouse Christian Setzkorn
2
Contents

• Description of the problem
• Background to trading rules
• Evolutionary approaches
• Aims
• Our implementation
• Results
• Conclusions

3
Problem description
Goal To produce a system that can efficiently
predict future prices of stocks
4
Background to trading rules

• Market prediction is a hard task (!) many
  variables play a role in market behaviour
• Over the years several technical indicators have
  been developed to describe stock performance
• Stochastic Oscillators
• Moving Average
• Momentum
• Bollinger Band
• EA techniques can cope fruitfully with market
  prediction, but.
• Literature concerning application of GP
  techniques is scarce and incomplete

5
Evolutionary approaches

• What are we evolving?
• Some sort of expert rule, a functional
  relationship between trading rules rule
  parameters.
• How will we evolve it?
• Grammatical evolution
• Genetic programming

6
Aims

• Compare a GP search engine against the existing
  GE.
• Modify the GE to include a new grammar and
  fitness function.
• Learn some stuff

7
Our approach

• Selection of EA technique
• Data pre-processing
• Choice of EA parameters
• Get results
• Interpret results

8
Our approach data pre-processing

• Selection of training and testing sets
• Normalisation

Test set 1 Training
set Test set 2
9
Our approach fitness evaluation

• Credit assignment
• Score for a correct decision
• Independent check for profit made

10
Results GP
11
Results GE

• Profits
• training set 840
• test set 2 -2992

12
Conclusions

• About our market timing system
• About our project
• What went wrong?
• What went well?
• Suggestions for future work

13
Our new system

• GP found a profitable solution but struggled to
  improve on it
• small population?
• inadequate fitness function?
• GE struggled to obtain a profitable solution
• same code problems?

14
What went wrong?

• GE system made small profit on training set but
• made a loss on the test set
• Little time for tuning
• No time for a thorough validation

15
What went well?

• First-attempt GP predictor made profits on
  training and test set.
• Increased awareness of EA methods.
• Good experience of a hectic implementation.

16
Future directions

• Validation and comparisons
• New financial indicators
• vertical horizontal filter, money flow index
• New fitness functions
• Sensitivity Specificity
• Consistency
• Complexity of solution
• Multiobjective approach
• Introduce different hold times
• Fuzzy rule system ? several rules