How Much Historical Data Do You Need?

Years Are a Proxy. Here’s the Real Spec.

The standard advice — FX Replay’s step-by-step guide prescribes 2–5+ years — points the right direction but measures the wrong thing. Two requirements actually determine sufficiency:

1. Regime coverage. The window must contain the market’s main weather systems: directional trends, dead ranges, and violent news-driven volatility. An edge measured in one regime is a one-regime edge — fine, if you know it; fatal, if you don’t.
2. Sample size. Whatever the calendar says, keep extending until the strategy has produced 100–200+ trades. A daily-chart system can need a decade to print 100 setups; a London scalp gets there in months.

By Strategy Type

Strategy type	Typical signal frequency	History that usually suffices
Scalping (M1–M5)	several/day	3–6 months — regimes cycle fast intraday
Intraday (M15–H1)	~1/day	6–12 months
Swing (H4–D1)	1–3/week	2–5 years
Position (D1–W1)	a few/month	5–10+ years, accepting older-data caveats

Recency vs Depth: Old Data Lies a Little

More history isn’t free accuracy, because markets drift: retail spreads have tightened over the years, volatility regimes change (compare 2020 to 2024 on any major), and microstructure evolves. Data from a decade ago tests your rules against costs and behavior that no longer exist. Practical weighting: validate on recent data first (the last 1–2 years carries the most predictive weight), use deeper history to stress the edge across regimes — not to fine-tune parameters against ancient conditions. Also check your source’s actual depth per instrument before planning (varies widely — replay tools publish theirs, e.g. StrategyTune’s catalog lists exact data ranges per instrument).

Reserve an Out-of-Sample Slice

The cheapest protection against manual overfitting: split the window. Develop and iterate on ~70–80% of it; keep the most recent 20–30% untouched until the rules are frozen, then run them once on the reserved slice. Matching expectancy → confidence. Collapsing expectancy → you tuned to noise, and the reserved slice just saved you the live-account version of that lesson. (For an honest read, the reserved slice also gets the full cost audit.)

Related: testing across regimes · sample size · free data sources

Frequently Asked Questions

Is one year of data enough to backtest?

For intraday strategies, usually yes — a year of M15 data spans hundreds of setups and several regime changes. For swing strategies it's marginal (perhaps 50-100 signals, possibly one dominant regime), and for position trading it's clearly insufficient. Let trade count and regime variety make the call.

Should I include 2020-style crisis periods?

Yes, deliberately — not because they'll repeat on schedule, but because they bound your worst case: crisis windows produce the drawdowns, gaps and spread blowouts that decide whether you'd survive holding the strategy. Segment results so crisis behavior is visible rather than averaged away.

What if my strategy didn't exist in older market conditions — do old results count against it?

Weight recent data more. If an edge works on the last two years but fails on data from a structurally different era (different spreads, volatility, participants), that's expected drift, not refutation. The honest framing: the strategy is validated for current conditions and needs monitoring as conditions evolve.

How do I do walk-forward testing manually?

A light version: split history into sequential blocks, develop on block one, validate on block two, then roll forward — develop on one+two, validate on three. It's more work than a single out-of-sample split, but for parameter-heavy strategies it shows whether re-tuning actually transfers, which a single split can't.