Backtesting with historical liquidation data: pitfalls and checks
How to test ideas on historical liquidation and heatmap data without fooling yourself: look-ahead bias, timestamps, coverage, gaps, baselines and costs.
On this page
- What liquidation history consists of
- Pitfall 1: look-ahead bias
- Pitfall 2: timestamps and delay
- Pitfall 3: coverage that changes under the test
- Pitfall 4: gaps
- Pitfall 5: rules chosen after the fact
- Pitfall 6: no baseline
- Pitfall 7: survivorship
- Pitfall 8: costs, if the test simulates trades
- Getting the data
- Common questions
A backtest asks how an idea would have behaved on past data. With liquidation data the idea is usually a question such as "how often did the price reach the largest liquidation cluster within a day?" or "what happened to funding after a large burst of long liquidations?". The answer is only as good as the data and the method. Most mistakes come from using information that was not known at the time, from coverage that changed under the test, and from choosing the rules after seeing the results.
This guide lists the pitfalls specific to liquidation and heatmap data and the checks that catch them. It is about research method, not about trading rules.
What liquidation history consists of
Three kinds of data are usually involved:
- Liquidation events. Positions that were actually liquidated, with time, coin, side, size and venue.
- Liquidation levels. Where open positions would be liquidated, as a map at one moment or a heatmap over time.
- Context series. Price candles, open interest, funding and long/short, used to describe what happened around the events.
Each has its own history and its own start date. The guide Crypto historical data explains which history exists for which data and why some of it can never be filled in afterwards.
Pitfall 1: look-ahead bias
Look-ahead bias means the test uses information that was not available at the moment it pretends to act on. With liquidation data this happens in quiet ways:
- Using today's map for yesterday. Liquidation levels change constantly. A test about a level must use the map as it stood at that time, not a later one.
- Revised data. A series that was corrected afterwards contains knowledge from after the fact. Point-in-time data is the version that was known at that moment.
- Daily values used intraday. A daily close or a daily total is only known at the end of the day. Using it to describe something at noon leaks the afternoon into the morning.
Our heatmap columns are stored snapshot by snapshot and never interpolated, and our daily archive files freeze each UTC day once without revising it later.1 A test that reads levels at time t from the snapshot taken at or before t sees only what was visible then.
Pitfall 2: timestamps and delay
Every record has at least two times: when something happened and when it became known. A liquidation happens at one moment, appears on a stream a moment later, and lands in a database later still. A heatmap snapshot reflects positions read over a few minutes.
Tests that compare events on a scale of seconds need to know these delays. Tests on a scale of hours are rarely affected. Either way, keep every series in UTC and align on the time the information was available, not only the time of the event.
Pitfall 3: coverage that changes under the test
A series can rise simply because more was being measured:
- Venues join. A liquidation total summed over venues jumps when a venue is added. Compare periods only where the same venues are covered.
- Tracked wallets change. Intraday heatmaps built from tracked positions depend on which wallets are polled. Large wallets are covered far better than small ones.
- Lower bounds. Exchange liquidation streams publish only part of the liquidations, so exchange totals undercount. Perp DEX vs CEX liquidations explains why.
A useful check is to rerun the test on one venue only, or on the daily snapshot of all positions, and see whether the result holds.
Pitfall 4: gaps
When a stream was silent or a collector did not run, the honest record is a gap. Filling it with the last known value or a smooth line creates observations nobody saw. Our gaps stay empty and are listed with their times on the historical data page.
In a test, exclude the affected windows and report how many observations were dropped. A result that depends on how gaps were filled is not a result.
Pitfall 5: rules chosen after the fact
The most common way to fool yourself is to try many definitions and keep the one that works. With liquidation data there are many knobs: how large a cluster must be, how far from the price, which time window counts as a hit, which coins are included.
Three habits help:
- Write the rules down first. Define the event, the window and what counts as a hit before looking at results.
- Split the data. Develop on one period and confirm on a later one that was not looked at.
- Report every variant you tried, not only the best one.
Our scoreboard works this way: each published cluster is recorded at the time with its size and distance, and whether the price reached it is checked later within fixed windows. Do liquidation clusters attract price? explains how to read that record.
Pitfall 6: no baseline
A hit rate on its own says little. Prices move, so any level near the price is reached fairly often. The question is whether liquidation clusters are reached more often than comparable levels without a cluster.
Distance and volatility explain most of how often a level is reached. A baseline at the same distance, in the same period, removes that effect.
Pitfall 7: survivorship
Coins are listed and delisted. A test that only uses coins trading today leaves out the ones that were removed, often after large moves and heavy liquidations. Keep delisted markets in the data for the periods they traded.
Pitfall 8: costs, if the test simulates trades
If a test simulates entering and exiting positions, it has to include what that would have cost: fees, funding paid or received while a position is open, and slippage when an order walks the book. Slippage grows with size and in fast markets, exactly when liquidations cluster. Order book depth and slippage explains the arithmetic. A result that disappears once costs are included was never there.
Getting the data
- CSV per coin. The History tab of each coin, for example BTC historical data, has daily rows and downloads within each plan's history.
- The API. Liquidations with cursor pages, heatmap history per coin and the scoreboard's hit rates are available over REST; see the API reference and Crypto liquidation data API.
- Daily archive files. Each UTC day frozen once, with a manifest and checksums for every file. A sample day is linked from the historical data page; the complete archive is available online as a dataset licence on Data pricing, for internal use within one organisation.
Exchange rows (Binance, Bybit) are in the API and the files as collected, without warranty, and exchange streams undercount; we may withhold a venue at any time, so a test that depends on single exchange events should also run on Hyperliquid data or on totals over venues.
Common questions
Can I backtest on a liquidation heatmap at all? Yes, if each test step uses the snapshot that existed at that time and the test accounts for which positions were tracked.
How far back does liquidation history go? It depends on the dataset and venue. The exact first day of each dataset is listed on the historical data page, because some start dates move back as more reconstructed history is imported.
Is a high hit rate evidence that clusters attract price? Only against a baseline at the same distance and period, with rules fixed in advance.
Sources
- Blockmetric, "Historical data" (accessed 2026-11-11)Back to text
How we collect and check the data: Data & methodology.
See today’s numbersThe live pages update continuously.
See what history exists