Parameter Optimization Guide

Parameter optimization is a core环节 in quantitative strategy development. Every strategy contains parameters that need tuning — the number of moving average periods, stop-loss ratios, screening thresholds, etc. Finding the right parameter settings can make or break a strategy.

But parameter optimization is a double-edged sword: used properly, it significantly improves strategy performance; used improperly, it causes severe overfitting.

What is Parameter Optimization?

The parameter optimization process:

1. Define Parameter Space: Define the range and step size for each parameter
2. Generate Candidate Combinations: List all parameter combinations to test
3. Run Backtests: Backtest each parameter combination
4. Evaluate Performance: Calculate performance metrics for each combination
5. Select Optimal Set: Choose the best parameters based on the objective function

A Concrete Example

Suppose your strategy logic is:

"Buy when the 20-day MA is above the 60-day MA; sell when the 20-day MA is below the 60-day MA."

Two parameters: 20 and 60. Parameter optimization tests different combinations (e.g., 10/30, 10/50, 20/60, 20/100, 50/200) to find the best-performing pair.

Common Parameter Optimization Methods

Method 1: Grid Search

Exhaustively test all possible parameter combinations:

• Parameter A: 5 options
• Parameter B: 5 options
• Parameter C: 5 options
• Total combinations: 5 × 5 × 5 = 125

Pros: Simple, thorough, guaranteed to find the best solution on the grid Cons: Computational cost grows exponentially with parameter count

Method 2: Random Search

Randomly draw N parameter combinations from the parameter space:

• Imagine 1 million possible combinations
• Randomly sample 1,000 for testing
• Select the best from results

Pros: Low computational cost, typically finds near-optimal parameters Cons: Cannot guarantee global optimum

Research finding: Bertrand and Bergstra (2012) showed random search typically outperforms grid search under equal computational resources.

Method 3: Bayesian Optimization

Intelligently select the next parameter set based on previous results:

1. Test an initial random parameter set
2. Update understanding of the parameter space based on results
3. Select the "most promising" next parameter set
4. Repeat until iteration limit reached

Pros: Finds good parameters with fewer tests Cons: Complex to implement, may be unstable on non-smooth parameter spaces

Method 4: Genetic Algorithm

Simulates natural selection:

1. Generate an initial parameter population
2. Evaluate fitness of each combination
3. Select the fittest for "crossover" (combination) and "mutation" (random adjustment)
4. Generate new generation
5. Repeat until convergence

Pros: Suits complex, non-linear parameter spaces Cons: Complex to implement, requires careful algorithm tuning

Best Practices for Parameter Optimization

1. Limit Parameter Count

Each additional parameter significantly increases overfitting risk:

• Conservative: Maximum 2-3 tunable parameters
• Moderate: Maximum 3-5 tunable parameters
• Complex: Maximum 5 tunable parameters

Rule: If a strategy needs more than 5 parameters to perform well, reconsider whether a simpler alternative exists.

2. Use Out-of-Sample Testing

The parameter optimization workflow should be:

Step 1: Optimize parameters on training set
Step 2: Validate optimal parameters on out-of-sample testing set
Step 3: If testing set performs well → proceed to Walk-Forward Analysis
Step 4: If testing set underperforms → revisit strategy logic

3. Focus on Parameter Stability

Good strategies have a "flat" parameter plateau:

• Parameters near the optimum should also perform well
• Minor parameter changes should not cause dramatic performance drops
• This indicates robustness rather than overfitting to specific values

4. Prioritize Economic Logic

Every parameter should have a clear economic rationale:

• 20-day MA: reflects short-term trend
• 60-day MA: reflects medium-term trend
• NOT: "I found 47 and 133 combine best in the data"

5. Avoid Multiple Testing Bias

If testing numerous parameter combinations simultaneously, use statistical corrections:

• Bonferroni correction: divide significance level by number of tests
• False Discovery Rate (FDR) control
• Or simplify strategy logic to reduce combinations

Common Parameter Optimization Traps

Trap 1: Optimizing on the Testing Set

This is the biggest mistake:

1. Optimize on training set
2. Validate on testing set
3. Testing set underperforms → return to training set to adjust
4. Repeat until testing set performance is "satisfactory"

This is equivalent to looking at exam answers before studying, then adjusting your review strategy to match.

Correct approach: Use the testing set only once. If it underperforms, do not re-optimize — redesign the strategy logic.

Trap 2: Chasing the Perfect Curve

Attempting to find a "perfect" backtest curve — smooth upward, almost no drawdown. This pursuit typically leads to severe overfitting.

Better goal: Find parameters that perform consistently across multiple market environments, not a single optimal curve.

Trap 3: Ignoring Transaction Costs

Using overly optimistic transaction cost assumptions during optimization:

• Ignoring slippage
• Ignoring bid-ask spreads
• Using unrealistically low commission rates

Correct approach: Use conservative transaction cost models including reasonable estimates for commissions, slippage, and spreads.

How Algo Lab Implements Parameter Optimization

Every Algo Lab strategy follows a strict parameter optimization workflow:

1. Parameter Count Limit: Maximum 5 tunable parameters per strategy
2. Random Search: Random search for initial parameter screening
3. Out-of-Sample Testing: 30% independent testing set validation
4. Walk-Forward Analysis: 5+ walk-forward windows validate parameter stability
5. Economic Logic Review: Each parameter must have a clear market mechanism explanation

Conclusion: Balancing Flexibility and Robustness

The core challenge of parameter optimization is balancing two seemingly contradictory goals:

• Flexibility: Parameters flexible enough to adapt to different market conditions
• Robustness: Parameters stable enough to avoid overfitting to specific historical patterns

Algo Lab's strategy development process achieves optimal balance through multi-layered validation mechanisms.

Learn about Algo Lab's strategy validation | Explore AI stock picking guide | [Join VIP for daily quantitative signals]

Frequently Asked Questions

What is Parameter Optimization?

Parameter optimization is the process of systematically testing different parameter combinations in a quantitative strategy to find the best settings for a given backtest period. For example, for a moving average strategy, optimization tests different MA periods (e.g., 10-day, 20-day, 50-day) to find the best combination.

Does parameter optimization always lead to overfitting?

Not necessarily. Parameter optimization itself is a legitimate strategy development process. The key is controlling overfitting risk: (1) limit parameter count; (2) validate with out-of-sample testing; (3) perform walk-forward analysis; (4) maintain economic logic for parameters. Following these principles, parameter optimization is a necessary step.

What method should I use for parameter optimization?

Common methods include: Grid Search, Random Search, and Bayesian Optimization. Grid search is simple but computationally expensive; random search is more efficient; Bayesian optimization suits large parameter spaces. For most strategies, random search offers the best cost-effectiveness.

How often should parameters be re-optimized?

Depends on strategy trading frequency and how quickly market conditions change. For daily-level strategies, quarterly re-evaluation is recommended. If walk-forward analysis shows good parameter stability, extend to bi-annual or annual re-evaluation.