Free Forex Candle Data (M1 OHLC)

28,412,683 one-minute OHLC bars across 15 instruments, 2021 to 2026. No account, no API key, no rate limit. This page also lists where the data is missing, which almost no free provider will tell you — and which matters more than the headline bar count.

What is available

Everything is stored at one-minute resolution and aggregated on demand to M5, M15, M30, H1, H4, D1 or W1 on true clock boundaries.

Where the data is missing

Every historical dataset has holes. Weekends and holidays are expected; unexplained multi-day gaps are not, and they will quietly distort any statistic computed across them.

These are the gaps longer than three days in this archive, which you can regenerate yourself:

node scripts/research.mjs holes GBPUSD

LAST BAR    NEXT BAR    DAYS
2021-01-31  2021-02-09  8
2021-03-10  2021-03-14  4
2021-10-13  2021-10-17  4
2023-12-29  2024-01-01  3
2025-08-21  2025-08-25  3
2026-03-12  2026-03-16  3

The eight-day gap in early 2021 is the largest in the set. If your study window covers February 2021, that absence is real and you should account for it rather than assume continuous coverage.

EURUSD carries a five-day gap around 2025-07-22, and several instruments share a hole in mid-March 2021 where the upstream source genuinely has no index data.

What the data does not contain

• No volume. Forex has no central exchange, so there is no true market-wide volume. Providers that report it are reporting their own feed's tick counts, which is a different thing wearing the same name. This archive omits the field rather than publish a number that invites false conclusions.
• No bid/ask spread. Bars are mid-price. Any strategy targeting a handful of pips will look better here than it would live, because the cost you would actually pay is absent.
• Not tick data. One minute is the finest resolution. For sub-minute execution studies you need a tick source.

How to get it

Three routes, depending on what you are doing.

1. Just use it in the browser

The free backtester at fxabsolute.com loads this archive directly. Nothing to download, no account.

2. Give it to a coding agent

The MCP server exposes the whole archive to Claude Code, Cursor or any MCP client, with aggregation and session filtering done server-side so you are not moving millions of rows around.

git clone --depth 1 --filter=blob:none --sparse https://github.com/varsansri/fxabsolute
cd fxabsolute && git sparse-checkout set mcp
cd mcp && npm install
claude mcp add fxabsolute -- node "$PWD/src/index.js"

3. Read the binary files directly

The .bin files are a deliberately trivial format — a 4-byte big-endian bar count, then fixed 20-byte records of uint32 timestamp and four float32 prices. Twenty bytes per bar instead of roughly sixty for the equivalent CSV row, which is why five years of M1 loads in a browser at all.

const dv = new DataView(buf)
const count = dv.getUint32(0, false)
for (let i = 0; i < count; i++) {
  const o = 4 + i * 20
  const time  = dv.getUint32(o,      false)
  const open  = dv.getFloat32(o + 4,  false)
  const high  = dv.getFloat32(o + 8,  false)
  const low   = dv.getFloat32(o + 12, false)
  const close = dv.getFloat32(o + 16, false)
}

How it compares to other free sources

Two honest notes on that table. Dukascopy is the upstream source for much of this archive — if you want raw ticks and are willing to do the assembly, go straight there. And Yahoo's roughly 30-day intraday window is the reason it is unsuitable for multi-year studies, not a criticism of the data itself.

What this archive saves you is the assembly step: fetching, de-duplicating, checking for empty responses and stitching years of chunks into something continuous. That work is genuinely tedious, and it is where most people's data projects quietly die.

When not to download at all

Downloading makes sense when you want the archive on disk — for a backtest engine, a database, or work that has to run offline. It is the wrong shape for two common cases.

Exploring. If you do not yet know which instrument or period you care about, pulling gigabytes to answer one question is wasteful. A query interface answers it in a call.

Feeding an AI agent. A model cannot usefully be handed a 38 MB CSV; it can only be handed what fits in its context. A tool call has no such ceiling, because the aggregation happens before the model sees a number.

For both, the same archive is queryable without downloading it:

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

The server pulls only the instrument you ask about, caches it under ~/.cache/fxabsolute-mcp, and never fetches the rest. So it is not really "no download" — it is a download you do not have to manage, scoped to what you actually used.

Frequently asked questions