How many years of data should a backtest cover?

Enough to include regimes that disagree with each other — at minimum one full drawdown-and-recovery cycle. A strategy tested only on 2023–2025 has never met a bear market; one tested through 2021–2026 has seen a euphoric run, a 25%+ index decline (2022), a recovery, and rate-regime whiplash. More history is generally better, with a caveat: markets change structurally (decimalization, ETF proliferation, zero commissions), so a rule's 1995 performance may describe a market that no longer exists. For daily-signal ETF strategies, five to ten years of minute-resolution data is a reasonable floor — and weight recent regimes more than distant ones when judging.