realjojo's picture
One-year edition: new dataset card; remove 4-coin order books and 19 Oct snapshot
3bbff04 verified
|
Raw History Blame Contribute Delete
19.9 kB
metadata
license: cc-by-4.0
pretty_name: Hyperliquid Market Data, One Year (Hydromancer Reservoir)
tags:
  - hyperliquid
  - crypto
  - perpetual-futures
  - liquidations
  - ohlcv
  - order-book
  - trades
  - defi
  - finance
  - time-series
  - parquet
task_categories:
  - time-series-forecasting
size_categories:
  - 1B<n<10B
configs:
  - config_name: liquidations
    default: true
    data_files: liquidations/*.parquet
  - config_name: candles_1m
    data_files: candles_1m/*.parquet
  - config_name: fills_perp
    data_files: fills_perp/*/*.parquet
  - config_name: fills_spot
    data_files: fills_spot/*.parquet
  - config_name: hip4_fills
    data_files: hip4_fills/*.parquet
  - config_name: twap_fills
    data_files: twap_fills/*.parquet
  - config_name: adl
    data_files: adl/*.parquet
  - config_name: orderbooks_perps
    data_files: orderbooks_perps/*.parquet
  - config_name: orderbooks_spot
    data_files: orderbooks_spot/*.parquet
  - config_name: orderbooks_outcome
    data_files: orderbooks_outcome/*.parquet
  - config_name: positions
    data_files: snapshots/positions_*.parquet
  - config_name: account_values
    data_files: snapshots/account_values_*.parquet
  - config_name: crash_fills
    data_files: crash_fills/*.parquet

Hyperliquid Market Data: One Year of Trades, Order Books, Liquidations and Positions

This dataset comes from Reservoir, Hydromancer's free archive of Hyperliquid market data.

It covers one full year, 18 November 2025 to 17 November 2026, for every market on every venue: Hyperliquid's own perps, all nine HIP-3 venues, spot pairs and HIP-4 prediction markets. Liquidations, ADL and HIP-4 fills go back further, to the start of Reservoir's history.

Part What it is Covers
fills_perp Every perp fill on every venue: Hyperliquid and the HIP-3 venues, one file per venue per day (about 199 GB) 18 Nov 2025 to 17 Nov 2026
fills_spot Every spot fill, one file per day (about 11 GB) 18 Nov 2025 to 17 Nov 2026
candles_1m 1-minute candles for every market: 463 perps on 10 venues, 336 spot pairs and 24,959 HIP-4 markets. 112,733,388 rows 18 Nov 2025 to 17 Nov 2026
orderbooks_perps Top 20 levels each side, every minute, for 463 perp markets on 10 venues. 149,049,153 snapshots 15 Dec 2025 to 17 Nov 2026
orderbooks_spot The same for 352 spot pairs. 139,531,655 snapshots 15 Dec 2025 to 17 Nov 2026
orderbooks_outcome The same for HIP-4 markets (75,326 YES and NO coins). 62,933,114 snapshots 2 May to 17 Nov 2026
liquidations Every liquidation on every venue: 7,862,966 fills from 529,041 accounts in 468 markets 28 Jul 2025 to 17 Nov 2026
adl Every auto-deleverage (ADL) on every venue: 44,728 fills on 111 days 26 Aug 2025 to 13 Nov 2026
twap_fills Every fill from a TWAP order, on every venue: 150,527,564 fills in 458 markets 18 Nov 2025 to 17 Nov 2026
hip4_fills Every fill on HIP-4 prediction markets since launch: 27,882,572 rows in 26,483 markets, covering 12,049,900 trades plus settlements, splits and merges 2 May to 17 Nov 2026
positions Every open perp position on every venue, on the 1st of each month, plus 15 Dec 2025 and 17 Nov 2026 (13 snapshots). The latest has 395,022 positions held by 132,227 accounts 15 Dec 2025 to 17 Nov 2026
account_values Every account's value at the same moments. The latest has 2,541,933 rows 15 Dec 2025 to 17 Nov 2026
crash_fills Every Hyperliquid perp fill on the day of the 10 October 2025 crash: 17,629,682 fills 10 Oct 2025

The full archive is updated every night on AWS and goes back to July 2025. See Get the full data below.

What you can do with it

  • Study liquidation cascades. See which markets, sizes and accounts were hit, minute by minute, during crashes like 10 October 2025.
  • Backtest a trading idea. Use a year of 1-minute candles, or every trade with fills_perp.
  • Compare venues. Every part covers Hyperliquid's own markets and every HIP-3 venue, side by side.
  • Research prediction markets. Every HIP-4 trade since launch, its order book every minute, and the YES or NO side on each row.
  • Measure liquidity. The order books show how much a trade of a given size would have moved the price, minute by minute, for a whole year.
  • Follow positioning over time. Monthly snapshots show every open position, its entry price, leverage and liquidation price.
  • Train or test models. Forecasting, risk and anomaly detection, on real exchange data with every trade recorded on-chain.

Load it

from datasets import load_dataset

liqs = load_dataset("Hydromancer/hyperliquid-reservoir", "liquidations", split="train")
candles = load_dataset("Hydromancer/hyperliquid-reservoir", "candles_1m", split="train")

fills_perp and the order books are large (about 199 GB and 23 GB). Read only the days or venues you need, straight from the Parquet files, with pandas or DuckDB:

import pandas as pd

liqs = pd.read_parquet("hf://datasets/Hydromancer/hyperliquid-reservoir/liquidations/")
xyz_day = pd.read_parquet(
    "hf://datasets/Hydromancer/hyperliquid-reservoir/fills_perp/xyz/fills_perp_xyz_2026-11-17.parquet"
)

Try it: the 10 October 2025 crash

The biggest day for liquidations is 10 October 2025: 394,452 liquidation fills, worth about $7.2 billion, across 41,287 accounts. To see which markets were hit hardest:

SELECT coin,
       count(*)                 AS fills,
       count(DISTINCT address)  AS accounts,
       round(sum(price * size)) AS usd_liquidated
FROM 'hf://datasets/Hydromancer/hyperliquid-reservoir/liquidations/liquidations_2025-10.parquet'
WHERE timestamp >= '2025-10-10 00:00+00' AND timestamp < '2025-10-11 00:00+00'
GROUP BY coin
ORDER BY usd_liquidated DESC
LIMIT 10;

What's in each part

fills_perp and fills_spot

Every fill, exactly as Reservoir stores it, one file per day:

  • fills_perp/<venue>/fills_perp_<venue>_YYYY-MM-DD.parquet, for example fills_perp/hyperliquid/fills_perp_hyperliquid_2026-11-17.parquet
  • fills_spot/fills_spot_YYYY-MM-DD.parquet

The columns are the same as liquidations below. Every trade appears twice, once for each account; filter on crossed = true to count each trade once. direction covers every kind of fill: Open Long, Close Short, Long > Short (a flip), the liquidation types, Auto-Deleveraging and Net Child Vaults. A venue only has files for the days it was trading (see Which folder is which venue).

Venue Days Size
hyperliquid 365 119.6 GB
xyz 365 63.9 GB
cash 164 8.8 GB
io 91 3.4 GB
km 157 1.1 GB
hyna 274 0.8 GB
para 231 0.6 GB
flx 215 0.4 GB
mkts 140 0.3 GB
vntl 214 0.2 GB
spot (fills_spot) 365 10.8 GB

Fills with a builder code are included; filter on the builder column to find them.

liquidations

One file per month, every venue together. Each row is one fill of a position that was liquidated: the liquidated account's side of the trade. Market liquidations have crossed = true. Backstop liquidations, where the liquidator vault takes over the position, have crossed = false, so don't filter this file on crossed. A single liquidation can be filled in several pieces, so it can span several rows that share the same tx_hash.

Main columns:

Column Meaning
timestamp Time of the fill (UTC)
coin Market, for example BTC or xyz:NVDA
dex Venue: hyperliquid, or a HIP-3 venue such as xyz
side buy or sell
direction Close Long or Close Short for most liquidations; Liquidated Cross Long, Liquidated Isolated Long (or Short) for the rest. Use LIKE '%Long' to catch all long liquidations
liquidation_method market (sold on the order book) or backstop (taken over by the liquidator vault; 68,564 fills)
price, size Fill price and size
liquidation_mark_px Mark price that triggered the liquidation
realized_pnl Profit or loss realised by this fill
start_position The account's position before this fill
fee, fee_token Fee paid
address The liquidated account
tx_hash, order_id, trade_id Identifiers for joining with other data

The other columns (builder, builder_fee, deployer_fee, priority_gas, twap_id and so on) follow Reservoir's fill format and are usually empty here. deployer_fee and priority_gas only exist from late March 2026.

candles_1m

One file per day. Built from Reservoir's 1-second candles, one row per market per minute that had trades. Minutes with no trades are missing, so check for gaps before running a backtest.

Column Meaning
minute Start of the minute (UTC)
coin Market name (see below)
dex Venue: hyperliquid, or a HIP-3 venue such as xyz
asset_class perp, spot or outcome (HIP-4)
open, high, low, close Prices for the minute
volume Volume traded in the minute

Each HIP-4 market has two coins, one for YES and one for NO.

How markets are named:

  • Hyperliquid perps: BTC
  • HIP-3 perps: venue prefix, for example xyz:MSFT
  • Spot: the pair's number, for example @107 (HYPE/USDC)
  • HIP-4 prediction markets: # and a number, for example #114940

twap_fills

One file per month, every venue together. Every fill that came from a TWAP order (an order split into small pieces over time). Same columns as liquidations; twap_id links the fills of one TWAP order.

hip4_fills

One file per month, hip4_fills_YYYY-MM.parquet. Same columns as liquidations, plus:

Column Meaning
outcome_id The HIP-4 market
outcome_side Which side of the market the coin is (YES or NO)

Prices run from 0 to 1, which you can read as the market's probability. direction is Buy or Sell for trades, and Split Outcome, Merge Outcome, Negate Outcome, Merge Question or Settlement for the other events. Every trade appears twice, once for each account, so to count trades and volume use crossed = true AND direction IN ('Buy', 'Sell'). Filtering on crossed alone also counts settlements, splits and merges, which roughly doubles the volume. HIP-4 markets started in USDH and later moved to USDC; the quote_symbol column still says USDH on every row, so don't rely on it.

adl

One file, adl_all.parquet, every venue together, with the same columns as liquidations. Each row is one fill of a position closed by auto-deleveraging. ADL is rare: most rows are from 10 October 2025.

orderbooks_perps, orderbooks_spot and orderbooks_outcome

One file per day for each market type, with one row per market per minute.

Column Meaning
coin Market, named as in candles_1m
dex Venue (perps only differ; spot and HIP-4 are all hyperliquid)
time Time of the snapshot (UTC)
block_time_ms, block_number Block the snapshot was taken at
bids, asks Top 20 price levels each side. Each level has px (price), sz (total size) and n (number of orders)

px and sz are stored as text to keep their exact value, so cast them to numbers before doing maths. Reservoir's order books start on 15 December 2025 (HIP-4 on 2 May 2026), so these parts start then. A market only appears on days it had a book; HIP-4 books include markets that never traded, so they list more coins than the candles.

positions and account_values

Snapshots of every account, taken a few minutes after midnight UTC on the 1st of each month from 1 January 2026, plus 15 December 2025 (the first snapshot in the window; Reservoir has none between mid-November and 14 December 2025) and 17 November 2026. Files are named positions_YYYY-MM-DD.parquet and account_values_YYYY-MM-DD.parquet; the snapshot's date is in the file name.

positions has one row per open perp position, on every venue:

Column Meaning
user Account
market Market, for example BTC or xyz:MSFT
size, notional Position size (negative = short) and its value in USD
entry_price, liquidation_price Average entry price and the price at which it would be liquidated
leverage_type, leverage Cross or isolated, and the leverage
funding_pnl Funding paid or received on the position
account_value The account's total value

account_values has one row per account and venue: account_value, total_long_notional, total_short_notional, account_mode and the portfolio-margin columns (is_unified, pm_ratio, pm_avail_by_branch, pm_total_account_value).

crash_fills

One file, fills_2025-10-10.parquet: every fill on Hyperliquid's own perp markets on 10 October 2025, from 89,831 accounts in 182 markets. It's before this year's window, so it's kept as its own part. Same columns as liquidations. Every trade appears twice, once for each account; filter on crossed = true to count ordinary trades once. Count liquidations with is_liquidation = true instead: in backstop liquidations the liquidated account is not the taker, and the liquidator vault's side has a direction starting with Liquidated. Counted this way, liquidations were $7.16 billion of the day's $34.8 billion (20.6%).

Everything else in Reservoir

Apart from the parts above, the full Reservoir archive on AWS has more, all as Parquet files, one folder per day (for example date=2026-10-08, in UTC): data from before 18 November 2025, every day after 17 November 2026, 1-second candles, daily snapshots including spot balances, and the raw fill feed.

Dataset Folder in s3://hydromancer-reservoir/ Starts Size for one day
Every fill on every venue (perps, spot, HIP-4) global/fills/raw/ 28 Jul 2025 about 640 MB
Perp fills, one venue by_dex/{venue}/fills/perp/all/ 28 Jul 2025 Hyperliquid about 360 MB
Liquidations, ADL, TWAP fills and builder fills by_dex/{venue}/fills/perp/liquidations/, adl/, twap_fills/, builder_fills/ 28 Jul 2025 (TWAP from 2 Aug 2025) liquidations about 1 MB
Spot fills global/fills/spot/all/ 28 Jul 2025 about 27 MB
HIP-4 prediction-market fills global/fills/outcome/all/ 2 May 2026 about 27 MB
1-second candles, every market global/candles/1s/ or by_dex/{venue}/candles/1s/ 31 Jul 2025 about 44 MB
Perp positions, spot balances, account values (daily snapshots) global/snapshots/perp/all/, spot/, account_values/ 29 Jul 2025 (some days missing) about 29 / 117 / 92 MB
Order books: top 20 levels each side, once a minute by_dex/{venue}/orderbook/1m/perps/ (Hyperliquid also has spot/ and outcome/) 15 Dec 2025 (HIP-4 from 2 May 2026) 0.2 to 0.4 MB per market

global/ has every venue in one place. by_dex/ splits the same data by venue, so the files are smaller when you only need one venue.

Which folder is which venue

Folder Venue Data from Data until
hyperliquid Hyperliquid's own perp markets 28 Jul 2025 ongoing
xyz Trade[XYZ] 13 Oct 2025 ongoing
flx Felix 13 Nov 2025 20 Jun 2026 (markets closed)
vntl Ventuals 13 Nov 2025 19 Jun 2026 (markets closed)
hyna HyENA 3 Dec 2025 2 Sep 2026 (markets closed)
km Markets.xyz (old folder) 12 Jan 2026 17 Jun 2026 (continues in mkts)
cash Dreamcash 20 Jan 2026 2 Jul 2026 (markets closed)
para Paragon 1 Apr 2026 ongoing
mkts Markets.xyz (current folder) 1 Jul 2026 ongoing
io Entropy 19 Aug 2026 ongoing

When new data appears

Reservoir is updated once a day. Each UTC day is added the next day, around 10:00 to 12:00 UTC. This Hugging Face copy is a dated snapshot and is not updated every day.

What Reservoir doesn't include

  • Today. The newest folder is the previous day. For today's data, use the Hydromancer API.
  • Every order. Order books are saved once a minute as the top 20 price levels. There's no order-by-order history; for that you need the live L4 stream.
  • Positions during the day. Snapshots are taken once a day, a few minutes after midnight UTC.
  • Hyperliquid's own markets before 28 July 2025, or a venue's data from before it launched.

Get the full data

The full archive lives in a public S3 bucket, s3://hydromancer-reservoir, in the Tokyo region (ap-northeast-1). The data is free. The bucket is "requester pays", so you need an AWS account, and AWS bills you for what you download.

Each day is added the following day, so the newest data is about a day behind.

Query a file in place with DuckDB, without downloading all of it. First put your AWS credentials in environment variables (the guide below shows how), then:

INSTALL httpfs; LOAD httpfs;
CREATE SECRET (TYPE s3, PROVIDER credential_chain, CHAIN 'env', REGION 'ap-northeast-1');
SET s3_requester_pays = true;

SELECT timestamp, coin, side, price, size, address
FROM read_parquet('s3://hydromancer-reservoir/by_dex/hyperliquid/fills/perp/liquidations/date=2026-10-18/fills.parquet')
WHERE coin = 'ETH'
ORDER BY timestamp
LIMIT 20;

Or download whole files with the AWS CLI:

aws s3 cp \
  s3://hydromancer-reservoir/by_dex/hyperliquid/fills/perp/liquidations/date=2026-10-18/fills.parquet \
  ./liquidations-2026-10-18.parquet \
  --request-payer requester

Downloading suits small files you'll use often, like liquidations. For big files, like a full day of fills, querying in place is usually cheaper, because you only transfer what you need.

The full guide, with a folder map, file sizes and more examples: How to Query Reservoir's Hyperliquid Data.

Reservoir docs: docs.hydromancer.xyz/reservoir. For data that Reservoir doesn't have, contact data@hydromancer.xyz.

For live data, Hydromancer has a Hyperliquid API and WebSocket streams.

FAQ

How is this different from Reservoir on AWS? This page holds one year (18 Nov 2025 to 17 Nov 2026) that you can open straight away, with no AWS account. Reservoir on AWS grows every night, goes back to July 2025 and also has 1-second candles, daily snapshots and the raw fill feed.

Is Reservoir real-time? No. It's updated once a day. For live data, use Hydromancer's WebSocket streams.

Do I need a Hydromancer API key for Reservoir? No. You need an AWS account, because the bucket is requester pays. An API key is only needed for the Hydromancer API.

Why is a date missing? A file only exists for a day that had matching data. For example, there are no ADL files on quiet days, and a HIP-3 venue only has files while it was trading. Some snapshot days are also missing in Reservoir.

How do I find a HIP-3 venue's data? Use the dex column, or the venue's folder in fills_perp/, such as fills_perp/xyz/ for Trade[XYZ]. See the venue table above.

What if the data I need isn't there? Email data@hydromancer.xyz.

Licence and citation

This data is licensed under CC BY 4.0. You can use, share and build on it, including for commercial work, as long as you credit Hydromancer and link to https://hydromancer.xyz.

Please cite as:

Hydromancer. Reservoir: Hyperliquid historical market data. https://docs.hydromancer.xyz/reservoir

About

Made by Hydromancer. The data is taken from Hyperliquid's public chain data.