CoEvolutionary Fuzzy Trading RuleBase Design Using Game Theory

In practice a Laplace smoothing of the payoff matrix avoids zerodivision problems, and a simple bestresponse update works well:

weight_i  weight_i  (1 + (avgPayoff_i  globalAvg))

where is a learning rate. The result is a selfregulating ecosystem that favours rule combinations which survive mutual competition.

4. Design Workflow

Step 1 Define Fuzzy Variables and Membership Functions

Choose the technical variables (price, MACD, Bollinger band width, etc.) and design linguistic terms (Low, Medium, High). Triangular or Gaussian membership functions are common because they are easy to differentiate for mutation operators.

Step 2 Encode Individuals

An individual encodes a complete fuzzy rule:

Gene 1 : Antecedent 1 (variable, term)Gene 2 : Antecedent 2 (variable, term)Gene 3 : Consequent (action, strength)Gene 4 : Operator (AND/OR)Gene 5 : Defuzzifier parameters

Step 3 Initialise SubPopulations

Create 35 species with random individuals, respecting each species semantic role (trend, reversal, risk). Keep population sizes modest (3050) to reduce computational load.

Step 4 Simulate Trading Episodes

Use historical OHLCV data segmented into windows (e.g., 250day rolling periods). For each window, draw one rule from each species according to its mixed strategy, combine the actions (e.g., weighted vote), and execute the simulated trade.

Step 5 Evaluate Payoffs

Common metrics:

• Sharpe ratio
• Maximum drawdown
• Profit factor

Combine them into a scalar payoff, for instance:

payoff = 0.5Sharpe  0.3Drawdown + 0.2ProfitFactor

Step 6 GameTheoretic Update

Construct payoff matrices for every pair of species, compute an approximate Nash equilibrium by iterating bestresponse dynamics, and use the resulting strategy probabilities to weight selection for the next generation.

Step 7 Genetic Operators

Apply crossover (swap antecedents between two parents) and mutation (perturb membership parameters, change operator, or replace a term). Preserve elite individuals (top 5%) to guarantee elitism.

Step 8 Termination and Extraction

Stop after a predefined number of generations (often 200300) or when improvement plateaus. The final rule base is assembled by taking the highestweight individual from each species and optionally pruning redundant rules using a correlation analysis.

5. Advantages of the Combined Approach

• Robustness: Coevolution prevents overfitting to a single market regime because species must succeed against each other.
• Interpretability: Fuzzy rules remain humanreadable, facilitating validation and regulatory compliance.
• adaptability: Gametheoretic pressure automatically shifts emphasis toward rule sets that perform well in the current environment.
• Scalability: Adding new species (e.g., newssentiment rules) integrates seamlessly as an extra player in the game.

6. Practical Considerations

Computation

Running full simulations for each generation can be expensive. Strategies to reduce load include:

• Parallelising the evaluation of individuals across CPU cores.
• Using a rollingwindow cache to reuse previously computed payoffs.
• Limiting the number of opponent samples per payoff calculation (MonteCarlo approximation).

Parameter Sensitivity

Key hyperparameters (population size, mutation rate, learning rate , payoff weighting) should be tuned via a separate metaoptimisation (grid search or Bayesian optimisation). Smallscale sanity checks on a single instrument help avoid runaway configurations.

Risk Management

Even though a riskmanagement species evolves separately, it is advisable to impose hard limits at the system level (e.g., maximum exposure per trade, portfoliowide VaR caps) to protect against pathological emergent behaviour.

7. Example Rule Set

Below is a concise illustration of a rule base that emerged after 250 generations on EUR/USD daily data.

1. IF price is SlightlyAboveMA(50) AND RSI is High THEN BUY strength 0.682. IF price is SlightlyBelowMA(200) AND MACD is Negative THEN SELL strength 0.553. IF ATR is High AND Volume is Low THEN REDUCE position by 30%4. IF BollingerWidth is Narrow AND StochK is Overbought THEN CLOSE LONG5. IF DailyReturnStd is Rising THEN DECREASE overall leverage by 10%

8. Concluding Remarks

Coevolutionary fuzzy rulebase design enriched with gametheoretic interactions offers a principled pathway to create adaptive, interpretable, and resilient trading systems. By treating rule species as strategic agents, the methodology captures the inherently competitive nature of financial markets while preserving the linguistic expressiveness of fuzzy logic. Future extensions may incorporate deep reinforcement learners as additional players, or employ evolutionary multiobjective optimisation to balance profitability against ecological stability.