Where to Get Free M1 Forex Data

One-minute forex history is genuinely free from several places — the differences are in how much work it takes to get it out and what breaks along the way. This page is the practical route list: access method, format, realistic effort, and the specific trap in each.

The routes, ranked by effort

Pick by what you are actually doing. If you want to practise, you do not need files at all — the first two rows put the data in front of you. If you are building a research pipeline or training something, you need real files, and HistData is the least painful source of those.

If you just want to practise: no download at all

Most people searching for M1 data want to backtest, not to own CSV files. Downloading and assembling gigabytes is a detour if that is your goal.

FXAbsolute has 28,412,683 one-minute bars across 15 forex, index and crypto instruments spanning 2021 to 2026 already loaded and replaying in the browser. No account, no download, no key. Bar-by-bar replay, drawing tools, limit orders, an automatic journal and live statistics.

Per-instrument depth, for reference:

HistData — the easiest real files

If you need actual files on disk, start here. Free, no account, M1 data back to 2000, delivered as monthly ZIP archives containing CSV.

What to expect: one ZIP per instrument per month. Five years of one pair is roughly sixty downloads, so scripting the fetch and the unzip is worth the ten minutes it takes. The CSV format is straightforward and consistent, which is more than can be said for API-based routes.

The traps: timestamps are in a specific timezone that you must confirm rather than assume, since silently mismatching it shifts every session boundary in your analysis. And concatenating months naively can leave duplicate or missing bars at boundaries — always de-duplicate on timestamp and then check bar counts per day.

Dukascopy — deepest, hardest

The deepest free source: tick data with bid and ask going back to roughly 2003, no account, no key, no cap. Most free forex archives, including ours, are built from it. Getting it out is genuinely difficult, and these are the specific failure modes rather than general warnings.

• Raise the connection timeout. The default in common client libraries is around ten seconds, which is too short for large requests. They fail on connect and look like missing data.
• Request roughly four-day chunks. Month-sized ranges reliably fail with connection resets; four-day windows reliably succeed.
• Retry empty responses, and never cache them. Dukascopy intermittently returns an empty array for a perfectly valid window — an empty success, not an error. Caching that as truth bakes a multi-day hole into your archive permanently, and it looks exactly like a weekend.
• Aggregate to M1 yourself. You receive ticks, so building minute bars is your job.

That third point is not theoretical. An audit of our own Dukascopy-derived archive found 7 coverage holes of 3 to 5 days across five years of EURUSD, six of them caused by exactly this behaviour. The full audit is published here.

Also upstream and unfixable: Dukascopy genuinely holds no index data for 10–18 March 2021, so index instruments carry an eight-day hole no amount of retrying will fill.

Two routes that will disappoint you

Alpha Vantage

The free tier is capped at 25 requests per day with an API key required. Downloading five years of one-minute history for a single instrument at that rate takes longer than anyone will persist with. It is a reasonable API for occasional quotes and the wrong tool for bulk history.

Yahoo Finance

Fast, free, easy through the yfinance library — and it retains only a short recent window of intraday data. You cannot pull years of M1 forex from it. Use it for daily bars, correlation work and quick prototypes, where it is genuinely good. Compared in detail here.

Verify whatever you download

Three checks, a few minutes, and they catch the failures that silently ruin backtests.

1. Count bars per trading day. A complete M1 day is 1,440 bars. Sort ascending; days far below that are partial, and zeros inside the trading week are holes.
2. Difference consecutive timestamps. Anything longer than a normal weekend needs an explanation before you test across it.
3. Confirm the timezone against a known event — a major release with a known publication time is ideal. Getting this wrong shifts every session-based conclusion you draw.

A quick sanity check on top: related pairs should stay related. Across 1,536 overlapping trading days, EURUSD and GBPUSD daily returns correlate at r = 0.77, moving the same direction 77.72% of the time. Sharp disagreement between two files on the same day means one of them is wrong.

If the data is for a script or an agent, skip the download

Everything above assumes you want files. If the data is destined for code — a notebook, a script, or an AI agent — there is a route that skips the download, the unzip and the CSV parsing entirely.

claude mcp add fxabsolute -- npx -y fxabsolute-mcp

That installs a published npm package which serves 28,412,683 one-minute bars across 15 instruments (2021 to 2026) to any MCP-capable client, with no key and no rate limit. Aggregation to M5 through W1 happens on request, on true clock boundaries.

Two properties matter more than convenience here. First, higher timeframes are bucketed on floor(ts / period) rather than by counting bars, so a weekend gap never smears an "H1 candle" across two sessions — a failure mode that silently corrupts a lot of homemade aggregation. Second, coverage is reported rather than assumed: ask for a range and the response tells you how many weekdays in it produced no data at all.

The archive carries no volume — a record is a timestamp plus four float32 OHLC values. If your strategy needs volume or VWAP, this is not your source.

Frequently asked questions