XAUUSD Backtesting Case Study: 4-Fold Walk-Forward Rerun

This XAUUSD backtesting case study shows public fold-level results from a SparklingAI gold research rerun completed on May 11, 2026 Malaysia time. The rerun used a larger execution dataset with 510,967 rows.

The goal is to document what the walk-forward test shows without exposing the proprietary alpha logic. The latest result is stronger than the previous snapshot because all four out-of-sample folds finished positive, but it should still be treated as research evidence rather than a live performance claim.

Walk-Forward Setup

The public result uses four out-of-sample folds:

• Fold 0: December 2021 to February 2023
• Fold 1: February 2023 to March 2024
• Fold 2: March 2024 to May 2025
• Fold 3: May 2025 to May 2026

Each fold starts with a $500 research equity baseline. The stitched walk-forward equity moved from $500.00 to $570.75 across all out-of-sample windows. The numbers below are aggregated public metrics from the test folds, not the private model recipe.

What The New Rerun Shows

The fold sequence is now positive across all four out-of-sample windows:

• Fold 0 ended around +5.50% with 20 trades, 70.0% win rate, and 2.50 profit factor.
• Fold 1 ended around +0.73% with 22 trades, 50.0% win rate, and 1.12 profit factor.
• Fold 2 ended around +5.87% with 11 trades, 81.8% win rate, and 8.05 profit factor.
• Fold 3 ended around +2.05% with 77 trades, 59.7% win rate, and 1.05 profit factor.
• The combined stitched out-of-sample result was around +14.15% after $24.01 in fees.

This is a useful improvement for the research track because the result is no longer carried by only one positive fold. The system showed positive behavior across several market windows.

Why Fold 3 Still Needs Careful Interpretation

Fold 3 is the most recent window and has the highest trade count, so it is valuable for studying current XAUUSD behavior. It also has the weakest profit factor and the highest drawdown in this rerun.

That means Fold 3 should not be presented as a simple victory. It is useful because it shows how the stack behaves in a more active recent period, but the low profit factor tells us the margin of error is still thin. This is exactly why SparklingAI tracks fold-level metrics instead of only reporting the final return.

Why Not Publish Every Diagnostic

The research folder contains more diagnostic experiments than the public chart shows. Some internal diagnostics test alternate profile heads, execution profiles, and research variants.

Those diagnostics are useful internally, but they are not the selected public walk-forward result. Publishing them in detail could overstate the result and reveal too much about the private alpha and execution stack. For public reporting, the cleaner approach is to show the selected fold-level metrics and explain the validation caveats.

What Is Not Being Published

This case study intentionally avoids publishing:

• Alpha construction
• Signal thresholds
• Feature engineering details
• Training configuration
• Exact execution rules
• Private model-selection logic

The public goal is to explain validation quality, not give away the system.

For the validation method behind this case study, read walk-forward testing for AI trading strategies. For the broader system direction, read what SparklingAI is building.