Directional changes presentation for EvoFIN 2015

1. Generating Directional Change Based
Trading Strategies with Genetic
Programming
Jeremie Gypteau
Fernando Otero
Michael Kampouridis

2. Background & motivation
• Majority of financial forecasting tools use a physical
time scale for studying price fluctuations
 Use snapshots of the market, taken at fixed intervals (e.g.
daily closing prices)
 This lacks realism
 Lose some significant activities
• Use intrinsic time scale
 Event-based approach
 Directional changes to model intrinsic time
 Genetic Programming to combine different directional
change strategies

3. Directional changes (DC)
• A DC event is identified by a change in the price of a
given stock
 Change is defined by a threshold value, decided by the
trader
 Upturn or downturn event
• After the confirmation of a DC event, an overshoot
(OS) event follows
 OS event finishes once an opposite DC event takes place

4. Directional changes
Source: Glattfelder, J., Dupuis, A., Olsen, R.: Patterns in high-frequency FX data: Discovery of 12 empirical
scaling laws. Quantitative Finance 11 (4), pp. 599-614 (2011)

5. Generating DC-based trading strategies
• A DC event is identified by a change in the price by
a given threshold value
• Different DC thresholds provide a different view of
data
 Smaller thresholds allow detection of more events and
hence actions can be taken promptly
 Larger thresholds detect fewer events, but provide the
opportunity of taking actions when bigger price variations
are observed
• Genetic Programming to combine the use of
different thresholds
 Automatically generate expressions that produce outputs
based on multiple threshold values

6. Genetic Programming configuration
Configuration Value
Individual structure tree
Function set Boolean functions {AND, OR, NOR,
XOR, NOT}
Terminal set Randomly generated boolean
terminals representing different DC
threshold values
Tree initialisation Ramped half-and-half
Genetic operators Subtree mutation, one-point
crossover, and reproduction
Selection Tournament selection
Termination criteria Maximum number of generations

7. Sample GP tree
Nor

8. GP Parameters
• Fitness = cash + (stockBalance * lastPrice)
Parameter Value
Max tree depth 8
Generations 300
Population 300
Reproduction 0.01
Crossover 0.97
Mutation 0.01
Elitism 0.01
Max DC threshold 10
Min DC threshold 0

9. Experiments
• 4 datasets
 2 stocks from FTSE 100 (Barclays, Marks&Spencer)
 2 international indices (NASDAQ, NYSE)
• Training: 1000 days, Testing: 500 days
• Aim
 Demonstrate that the paradigm of DC returns profitable
strategies
 Provide evidence that the strategies generated by the GP
are more profitable than using a fixed threshold
• Fixed thresholds: 0.02, 0.05, 0.10, 0.20, 0.50, 1.0, 1.50,
2.00, 2.50, 3.00, 4.00, 4.50, 5.00

10. Test results
Dataset GP Fixed DC
Barclays +4.67% -14.07%
Marks & Spencer +0.55% -0.33%
NASDAQ +4.93% +1.78%
NYSE +7.31% +4.80

11. Summary
• Intrinsic vs physical time
• DC as an alternative
• Positive results
 Positive returns in all 4 stocks tested
 Not always the case for Fixed DC results
 Overall returns for GP+DC >> Fixed DC
• Future research directions
 Explore different parameters settings, e.g. optimise the
trading amount
 Move to FX datasets, where several scaling laws have
been observed, and could thus increase the profit margin

12. Questions?
• Please email Michael Kampouridis at:
 M.Kampouridis ([at] kent [dot] ac [dot] uk)