Read the indicator's sign and distribution, avoid two common misreads, and put it into a strategy.
After a sharp drop, Liquidation History shows a deep negative bar. Were a lot of longs liquidated, or does this coin's value usually sit below 0? The indicator reads only Binance, and 0 is not a balance point; those two facts decide how you read that bar. If you are looking at liquidations by price level, that is a separate model estimate: see How to Read the Liquidation Map.
The liquidation indicator is a scaled value that shows whether shorts or longs were liquidated more over a period, built only from Binance's public liquidation feed. A liquidation is the exchange force-closing a leveraged position: a liquidated long becomes a sell order, a liquidated short a buy order. Every 5 minutes Blave sums each side's liquidation quantity (in coins) per symbol, then computes the indicator in three steps:
| Step | What it computes |
|---|---|
| 1. Net liquidations | Every 5 minutes, short liquidations minus long liquidations, in coins, not USD. A 5-minute slot with no liquidation record counts as 0. |
| 2. Rolling sum | Summed over the chosen timeframe (default 24h). |
| 3. Scaling | Divided by the 30-day rolling standard deviation of that sum. The mean is not subtracted, so it is not a standard z-score. |
The value has no bounds. Its sign tells you which side was liquidated more:
| Value | Meaning |
|---|---|
| > 0 | More shorts than longs were liquidated (in coins) over the timeframe |
| < 0 | More longs than shorts were liquidated |
| ≈ 0 | The two sides offset, there were no liquidation records, or the feed was down. It does not mean the coin's longs and shorts are balanced |
The distribution is asymmetric: the lower tail (long liquidations) runs deeper than the upper tail. These quantiles are Blave's own computation from the Binance force-order feed, on 1h bars with a 24h timeframe, 2024-09-15 to 2026-09-14, one value per hour (n = 17,520):
| Coin | p1 | p50 | p99 | Share of time above 0 |
|---|---|---|---|---|
| BTC | −3.76 | −0.01 | +3.31 | 48.8% |
| ETH | −4.07 | −0.08 | +2.79 | 43.1% |
| SOL | −4.01 | −0.17 | +2.87 | 37.6% |
| DOGE | −4.07 | −0.23 | +2.97 | 31.1% |
Liquidation Overview, History and the 24h bars on the map page read only the Binance force-order feed. Binance pushes at most one liquidation per symbol per second and drops the rest in that second, so this is a sample that undercounts in big moves.
Units are coins, so raw liquidation amounts cannot be compared across coins; compare the scaled value instead. While the feed is down, the value looks as if nothing was liquidated, and the gap is not flagged.
Above 0 means more shorts than longs were liquidated; below 0, more longs. But the mean is not subtracted when scaling, so a coin where longs get liquidated more over time drifts negative, and 0 is not that coin's normal level.
The table above gives the share of time above 0 over that measurement period (1h bars, 24h timeframe). The same 0 sits near BTC's median, but for DOGE it already leans further toward short liquidations than most of the time.
No official strategy uses liquidation data today, so all three uses below are sketches, not backtested. The thresholds in the code borrow the quantiles measured above as a starting point. They did not come from a parameter scan, so backtest before you use them.
The most direct approach flags the periods when the value lands in either tail. Because the lower tail runs deeper, do not use the same absolute value for both sides. For BTC, p1 is about −3.8 and p99 about +3.3.
LONG_FLUSH_TH = -3.8 # sketch: about BTC's p1 at 1h × 24h SHORT_FLUSH_TH = 3.3 # sketch: about BTC's p99 at 1h × 24h liq = fetch_liquidation(SYMBOL, INTERVAL, START, END, hdrs, timeframe='24h') df['LIQ'] = liq['alpha'] long_flush = df['LIQ'] <= LONG_FLUSH_TH # concentrated long liquidations short_flush = df['LIQ'] >= SHORT_FLUSH_TH # concentrated short liquidations
How to trade the flags is your research: Blave has not verified whether price reverses or continues after concentrated liquidations. Quantiles also differ by coin. ETH, SOL and DOGE all have a deeper p1 than BTC, so recompute when you switch coins.
Since 0 is not a balance point, a rule like "above 0 means shorts are getting liquidated" breaks down on coins like DOGE that lean negative for long stretches. Use the coin's own quantiles over a past window instead, which is less affected by that long-run lean:
# sketch: quantiles over the past 180 days; window length not scanned
lo = df['LIQ'].rolling('180D').quantile(0.01)
hi = df['LIQ'].rolling('180D').quantile(0.99)
extreme = (df['LIQ'] <= lo) | (df['LIQ'] >= hi)
A window that is too short lets the thresholds drift with recent price action. Also keep the window out of the warm-up period, roughly the first 30 days of a coin's data (before 2023-01-31 for BTC and ETH), when the standard deviation had too few samples and produces false extremes.
A more conservative use adds it to an existing strategy's entry rule: skip new positions while the value sits in either tail. It reuses the two thresholds from use 1. Replace compute_entry with your strategy's own entry condition:
# sketch: no entry while the value sits in either tail flush = (df['LIQ'] <= LONG_FLUSH_TH) | (df['LIQ'] >= SHORT_FLUSH_TH) entry = compute_entry(df) & ~flush
This combination is not backtested, and Blave has not verified that avoiding concentrated liquidations improves performance. It only treats "a wave just got liquidated" as a reason to pause new entries.
The indicator has two settings. The timeframe is the window for summing net liquidations: 15min, 1h, 4h, 8h, 24h or 3d, default 24h. The bar interval (period) sets how often a value is taken: 5min, 15min, 1h, 4h, 8h or 1d. Values are always computed on the 5-minute grid first, then each bar takes the last one. Changing the timeframe changes the distribution; here is the same period measured with a 1h timeframe on 1h bars:
| Coin | p1 | p50 | p99 | Share of time above 0 |
|---|---|---|---|---|
| BTC | −2.62 | 0.00 | +2.12 | 52.5% |
| ETH | −2.84 | 0.00 | +2.17 | 48.8% |
| SOL | −3.07 | 0.00 | +2.21 | 45.5% |
| DOGE | −3.50 | 0.00 | +2.09 | 40.3% |
Compared with 24h, 1h has narrower tails, a median of 0.00 for every coin, and a share above 0 closer to half. So recompute thresholds whenever you change the timeframe; one set of numbers does not carry over. 15min, 4h, 8h and 3d were not measured here.
| Page | What it shows | Access |
|---|---|---|
| Liquidation Overview | A treemap with a dropdown for short-liquidation and long-liquidation rankings, top 30 each. Tile size is the absolute value (relative to the coin's past 30 days), not a dollar amount; color is the price change over the timeframe. | Log in |
| Liquidation History | Bars of one coin's liquidation indicator over price, with timeframe and bar interval options and the stat fields. | Last 7 days delayed for non-Pro |
These are ideas for combinations, not verified conclusions:
| Pair | Use |
|---|---|
| Liquidation + Whale Hunter | Use Whale Hunter to spot unusual open-interest moves, then check which tail the liquidation indicator sat in over the same period to see whether the move came with concentrated liquidations. |
| Liquidation + Taker Intensity | The liquidation indicator only sees the side that was forced out. Use Taker Intensity to see aggressive buying and selling over the same period and cover the voluntary half. |
The indicator endpoint is GET /liquidation/get_alpha, with symbol, period, timeframe (default 24h) and a date range (start_date, end_date); in Blave Agent it maps to fetch_liquidation. It returns closed bars only, with timestamp, alpha and stat, and ranges over 1 year are truncated. Data starts 2023-01-01 for the earliest symbols (BTC, ETH), with earlier values at 0; other coins may start later. The 30-day standard deviation needs a warm-up, so meaningful values start about 30 days after a coin's data begins (around 2023-01-31 for BTC and ETH).
The stat object from get_alpha describes the latest value, and its fields do not mean quite what their names suggest:
| Field | What it actually computes |
|---|---|
up_prob | A logistic regression trained on the past 365 days uses the liquidation indicator to predict whether price is up or down 24h later, and returns the probability of up. The training data includes the current period, so it is an in-sample estimate, not an out-of-sample win rate. |
avg_up_return /avg_down_return | Average 24h return across all up / down outcomes in the past 365 days, regardless of the indicator's level. |
exp_value | up_prob × avg_up_return + (1 − up_prob) × avg_down_return, as a decimal, not a percentage. |
is_data_sufficient | Whether the coin's data starts more than 365 days ago, not whether this reading level has enough samples. |
For the Liquidation Map, Map Change and By Exchange endpoints and fields, see How to Read the Liquidation Map.
stat is an in-sample estimate. Verify any threshold with your own backtest.
