Beginner12–18 min

Lesson 7. Optimization Without Overfitting — Safe Ways to Tune Parameters

Learn grid vs genetic search, build a robust objective (PF with drawdown/stability penalties), split IS/OOS, plan walk-forward, and stress-test with costs/volatility. Spot red flags of overfitting before they bite live.

You ran realistic backtests. Now tune parameters safely and validate robustness.

7.1. Parameter search: Grid vs Genetic (when and why)

Two common ways to scan parameter space: brute-force grids and genetic (heuristic) search.

Grid Search

How it works:

Try every combination of parameter values in the defined ranges.

Pros:

Exhaustive — tests all combinations
Simple to reason about
Reproducible results

Cons:

Explodes with more parameters/steps
Very slow for large spaces

Use when: ≤ 10k combinations; few parameters (2–3); quick sanity checks.

Genetic Search

How it works:

Heuristic evolution toward better regions of parameter space.

Pros:

Faster on big spaces
Finds good-enough sets quickly
Handles many parameters

Cons:

Non-deterministic (results vary)
Can miss narrow peaks
Must control overfitting via validation

Use when: > 10k combinations; 3–6+ parameters with wide ranges.

If grid size > 50k, consider genetic or shrink ranges.

7.2. Parameter Space Estimator (how big is your grid?)

Before running optimization, estimate how many combinations you'll test.

Param 1

Min

Max

Step

Steps: 26

Param 2

Min

Max

Step

Steps: 17

Param 3

Min

Max

Step

Steps: 16

Toggle

Min

Max

Step

Steps: 2

Avg evaluation time (sec)

Combinations: 14,144

Estimated runtime: 1.2 hours

10k–50k → Consider genetic search

7.3. Objective function: PF with penalties for DD/Stability

Optimize for robustness, not just PF peaks.

PF (Profit Factor)

Gross Profit / Gross Loss

Max DD (relative)

Peak-to-trough drawdown as % of equity

Stability

E.g., percent positive months or equity slope × R² proxy

Build Your Objective

Score = PF − α·(DD/20) − β·(1 − Stability)

Drawdown penalty (α)0.5

Instability penalty (β)0.5

We'll prefer parameter sets with PF≥1.2, DD≤15%, and stable equity. Score penalizes DD (weight 0.5) and penalizes instability (weight 0.5).

7.4. IS/OOS and Walk-Forward planning

IS/OOS Split

Split your data into In-Sample (IS) for optimization and Out-of-Sample (OOS) for validation.

In-Sample (months)

Out-of-Sample (months)

Split: IS 12m → OOS 3m

Walk-Forward

Walk-forward: optimize on IS, lock parameters, evaluate on next OOS, repeat.

Number of folds

Schedule: 4 folds: IS 12m → OOS 3m ×4

Total data needed: 60 months

Pass rules: PF ≥ 1.2 on OOS & Max DD ≤ 15% in each fold

Optimize on IS, lock parameters, evaluate on the next OOS; repeat.

7.5. Stress testing: costs and volatility scenarios

A robust parameter set survives realistic cost and volatility changes.

ATR percentile < 30

Spread × / Slip +	+0 pip	+0.2 pip	+0.5 pip
×1
×1.2
×1.5

Robustness: Passes 0 of 9 cost cells in Low regime

Note: Record which scenarios fail and why (slippage/vol spikes).

7.6. Overfitting red flags (quick diagnostic)

PF on IS

PF on OOS

Parameter count

Grid size (combinations)

Stability (% positive months)

Sensitivity drop (% PF change ±10% params)

Overfitting Risk: Medium

PF_OOS / PF_IS = 0.72

Guidance: Cut parameters; use penalties; increase OOS; re-run WF; re-check costs/volatility.

High risk flags:

PF_OOS / PF_IS < 0.7
OOS < 3 months
Parameter count > 7
Grid > 50k with no walk-forward
Sensitivity drop > 20%

Lesson 7 Quiz

Test your understanding with 3 questions. Pass with 2/3 correct.

7.9 What's Next?

You've learned to optimize parameters safely and validate with walk-forward and stress tests. Next, we'll cover execution: ping, slippage, VPS — the infrastructure that keeps your robot running smoothly.

Educational content only. Not financial advice. Trading involves risk of capital loss. Backtests and past performance do not guarantee future results.

7.1. Parameter search: Grid vs Genetic (when and why)

Two common ways to scan parameter space: brute-force grids and genetic (heuristic) search.

Grid Search

How it works:

Try every combination of parameter values in the defined ranges.

Pros:

Exhaustive — tests all combinations
Simple to reason about
Reproducible results

Cons:

Explodes with more parameters/steps
Very slow for large spaces

Use when: ≤ 10k combinations; few parameters (2–3); quick sanity checks.

Genetic Search

How it works:

Heuristic evolution toward better regions of parameter space.

Pros:

Faster on big spaces
Finds good-enough sets quickly
Handles many parameters

Cons:

Non-deterministic (results vary)
Can miss narrow peaks
Must control overfitting via validation

Use when: > 10k combinations; 3–6+ parameters with wide ranges.

If grid size > 50k, consider genetic or shrink ranges.

7.3. Objective function: PF with penalties for DD/Stability

Optimize for robustness, not just PF peaks.

PF (Profit Factor)

Gross Profit / Gross Loss

Max DD (relative)

Peak-to-trough drawdown as % of equity

Stability

E.g., percent positive months or equity slope × R² proxy

Build Your Objective

Score = PF − α·(DD/20) − β·(1 − Stability)

Drawdown penalty (α)0.5

Instability penalty (β)0.5

We'll prefer parameter sets with PF≥1.2, DD≤15%, and stable equity. Score penalizes DD (weight 0.5) and penalizes instability (weight 0.5).

7.4. IS/OOS and Walk-Forward planning

IS/OOS Split

Split your data into In-Sample (IS) for optimization and Out-of-Sample (OOS) for validation.

In-Sample (months)

Out-of-Sample (months)

Split: IS 12m → OOS 3m

Walk-Forward

Walk-forward: optimize on IS, lock parameters, evaluate on next OOS, repeat.

Number of folds

Schedule: 4 folds: IS 12m → OOS 3m ×4

Total data needed: 60 months

Pass rules: PF ≥ 1.2 on OOS & Max DD ≤ 15% in each fold

Optimize on IS, lock parameters, evaluate on the next OOS; repeat.

Spread × / Slip +

+0 pip

+0.2 pip

+0.5 pip

×1

×1.2

×1.5

7.6. Overfitting red flags (quick diagnostic)

PF on IS

PF on OOS

Parameter count

Grid size (combinations)

Stability (% positive months)

Sensitivity drop (% PF change ±10% params)

Overfitting Risk: Medium

PF_OOS / PF_IS = 0.72

Guidance: Cut parameters; use penalties; increase OOS; re-run WF; re-check costs/volatility.

High risk flags:

PF_OOS / PF_IS < 0.7
OOS < 3 months
Parameter count > 7
Grid > 50k with no walk-forward
Sensitivity drop > 20%

Lesson 7. Optimization Without Overfitting — Safe Ways to Tune Parameters

7.1. Parameter search: Grid vs Genetic (when and why)

7.2. Parameter Space Estimator (how big is your grid?)

7.3. Objective function: PF with penalties for DD/Stability

7.4. IS/OOS and Walk-Forward planning

7.5. Stress testing: costs and volatility scenarios

7.6. Overfitting red flags (quick diagnostic)

7.7. Practice: run a small, safe optimization and validate (10–20 min)

Lesson 7 Quiz

7.9 What's Next?

Lesson 7. Optimization Without Overfitting — Safe Ways to Tune Parameters

7.1. Parameter search: Grid vs Genetic (when and why)

7.2. Parameter Space Estimator (how big is your grid?)

7.3. Objective function: PF with penalties for DD/Stability

7.4. IS/OOS and Walk-Forward planning

7.5. Stress testing: costs and volatility scenarios

7.6. Overfitting red flags (quick diagnostic)

7.7. Practice: run a small, safe optimization and validate (10–20 min)

Lesson 7 Quiz

7.9 What's Next?