Datasets:
|
Download README.md from ironflow/hyperliquid-trades: direct link, hf CLI and curl.
- Browser
- Download file 7.08 kB
-
https://huggingface.co/datasets/ironflow/hyperliquid-trades/resolve/main/README.md
- Command line
-
hf download hf://datasets/ironflow/hyperliquid-trades/README.md
-
curl -L -o README.md https://huggingface.co/datasets/ironflow/hyperliquid-trades/resolve/main/README.md
7.08 kB
| license: cc-by-nc-4.0 | |
| pretty_name: Hyperliquid Trades | |
| task_categories: | |
| - time-series-forecasting | |
| tags: | |
| - hyperliquid | |
| - crypto | |
| - perpetual-futures | |
| - defi | |
| - trades | |
| - tick-data | |
| - market-microstructure | |
| - spot | |
| - finance | |
| - trading | |
| - time-series | |
| size_categories: | |
| - 100M<n<1B | |
| configs: | |
| - config_name: default | |
| data_files: | |
| - split: train | |
| path: "data/*.parquet" | |
| # Hyperliquid Trades | |
| **Publishing from 10 November 2026:** new files appear 90 days after the data date, carry no wallet addresses, and are licensed CC BY-NC 4.0 (non-commercial). Files published before that date stay under CC BY 4.0. For recent data, wallet-level data or commercial use: https://ironflow.sh/?utm_source=huggingface | |
| Every trade on Hyperliquid perps and spot from 3 October 2026, one row per trade with the buyer, the seller and the taker side, with a new file every day. <!-- coverage -->Through 8 November 2026: 37 days, 210,257,579 trades.<!-- /coverage --> | |
| In the first 30 days, 3 October to 1 November 2026: 167,577,425 trades across 641 markets (179 native perps, 171 HIP-3 builder perps on `xyz`, `para`, `io` and `mkts`, and 291 spot pairs), $233.6 billion in notional between 327,481 wallets. | |
| BTC-PERP is the largest market at $68.1 billion; HYPE-PERP has the most trades, 9.4 million. HIP-3 markets took 31% of the notional, $73.1 billion, almost all of it on `xyz`. The busiest day is 29 October 2026, with 10.2 million trades and $18.2 billion; its 20:00 UTC hour alone had 1.45 million trades and $3.6 billion. | |
| Each row joins the two fills of one trade, so the buyer, the seller and the taker are on one line. The fills are the ones [Ironflow](https://ironflow.sh/?utm_source=huggingface) records from its own Hyperliquid node, the same data its `GET /v1/trades` API serves one fill at a time. | |
| ## Get live data | |
| Live Hyperliquid data is on [Ironflow](https://ironflow.sh/?utm_source=huggingface). The same trades, up to the latest block, are on Ironflow's API: [`GET /v1/trades`](https://docs.ironflow.sh/api-reference/market-data/list-trades), one row per fill (both sides of each trade). No key is needed for the last 24 hours; a [free key](https://ironflow.sh/key) gives 14 days. | |
| ## Source and updates | |
| - **Source:** fills recorded by Ironflow's own Hyperliquid node. Block gaps, such as node outages, are refilled automatically from Hyperliquid's node archive on S3. | |
| - **Checks:** row counts match Ironflow's store day by day, hourly trade counts match Hyperliquid's own candles, and a sample matches Ironflow's API field by field (see Notes). | |
| - **Updates:** one file per UTC day, added daily 90 days after the data date; the coverage line at the top is updated with it. | |
| - **Coverage:** from 3 October 2026. Native perps, HIP-3 perps and spot; HIP-4 outcome markets are not included. | |
| ## Files | |
| One Parquet file per UTC day: `data/YYYY-MM-DD.parquet`, one row per trade, in exchange order (time, then block, then position in the block). A day holds 3 to 10 million trades, 80 to 250 MB. | |
| | Column | Type | Meaning | | |
| |---|---|---| | |
| | `timestamp` | timestamp (ms, UTC) | trade time | | |
| | `market` | string | Ironflow market symbol, e.g. `BTC-PERP`, `xyz:NVDA-PERP`, `HYPE/USDC` | | |
| | `coin` | string | Hyperliquid's name for the market, e.g. `BTC`, `xyz:NVDA`, `@107` | | |
| | `dex` | string | HIP-3 builder dex (`xyz`, `para`, ...); empty for native perps and spot | | |
| | `market_class` | string | `perp` or `spot` | | |
| | `price` | float | trade price | | |
| | `size` | float | trade size in the base asset | | |
| | `notional` | float | `price * size`, in the market's quote currency (see Notes) | | |
| | `taker_side` | string | `buy` if the taker bought, `sell` if the taker sold | | |
| | `buyer` | string | address of the buying wallet | | |
| | `seller` | string | address of the selling wallet | | |
| | `liquidation` | bool | true when a fill of the trade carries Hyperliquid's `liquidation` field | | |
| | `block` | int | Hyperliquid block | | |
| | `tid` | int | Hyperliquid trade id, shared by the buy fill and the sell fill | | |
| The taker is the buyer when `taker_side` is `buy` and the seller otherwise. | |
| ## Use | |
| ```python | |
| import pandas as pd | |
| t = pd.read_parquet("hf://datasets/ironflow/hyperliquid-trades/data/2026-10-29.parquet") | |
| btc = t[t.market == "BTC-PERP"].set_index("timestamp") | |
| flow = btc.notional.where(btc.taker_side == "buy", -btc.notional).resample("5min").sum() | |
| print(flow.tail()) # taker net buying on BTC, USDC per 5 minutes | |
| ``` | |
| ```sql | |
| -- DuckDB: the 10 busiest markets and the share of volume bought by takers | |
| SELECT market, sum(notional) / 1e9 AS usd_bn, count(*) AS trades, | |
| sum(notional) FILTER (WHERE taker_side = 'buy') / sum(notional) AS taker_buy_share | |
| FROM 'data/*.parquet' | |
| GROUP BY market | |
| ORDER BY usd_bn DESC | |
| LIMIT 10; | |
| ``` | |
| ## Notes | |
| - **One row per trade.** Hyperliquid records a fill for each wallet in a trade, and Ironflow's API returns both. Here the buy fill and the sell fill that share a `tid` are joined; the taker is the fill Hyperliquid marks as crossing. | |
| - **Not included:** HIP-4 outcome markets, where many fills have one side only and do not pair into trades; and fills with trade id 0, which are spot sells worth under $1 each at 00:00 UTC every day, with no buyer. | |
| - **Checks.** Each day's row count equals the number of taker fills in Ironflow's store. Over all 720 hours, hourly trade counts equal Hyperliquid's own `candleSnapshot` in 5,024 of 5,040 market-hours for BTC, ETH, HYPE, SOL, `xyz:XYZ100` and `xyz:SKHX` perps and HYPE/USDC; the other 16 have 1 to 6 more trades here and none has fewer. In the HYPE/USDC hours the extra trades are exactly the liquidation trades, which Hyperliquid's candles leave out. On 1 November 2026, 2,259 trades in BTC-PERP, HYPE/USDC and `xyz:XYZ100-PERP` windows match Ironflow's `GET /v1/trades` in time, price, size, buyer and seller. | |
| - Ironflow's store holds a second copy, without a block number, of 30 `para:PAPER-PERP` trades on 19 October 2026 (21:39 to 22:01 UTC); only the copy with the block is kept. | |
| - `notional` is in the market's quote currency, a USD stablecoin. For spot it is the token after the slash: USDC for 286 of the 291 pairs, USDT0 or USDE for the other 5. | |
| - There is no transaction hash column; `block` and `tid` identify a trade. | |
| - `liquidation` is true for 611,864 trades. For one row per liquidated position, see [ironflow/hyperliquid-liquidations](https://huggingface.co/datasets/ironflow/hyperliquid-liquidations). | |
| - 1,275 spot trades, a few dozen a day at 00:00:00 UTC, have `0xeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee` as one side. | |
| - No duplicate (market, `tid`) rows and no trades with the same wallet on both sides. | |
| - Wallet addresses are public on-chain data. To object to, or ask us to remove, a wallet address, write to hello@ironflow.sh. Files already published cannot be recalled; later files leave the address out. | |
| ## Changes | |
| - **v1, 2 November 2026.** First release: 3 October to 1 November 2026. | |
| ## License and citation | |
| [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Credit "Ironflow (ironflow.sh)" when you use or share the data. | |