Does the model beat the baseline?
Last 30 complete days · 2026-08-28 to 2026-09-27 UTC- insufficient data4-hour horizon: not enough data for a verdict yet (0 of the 48 scored snapshots needed).
- insufficient data24-hour horizon: not enough data for a verdict yet (0 of the 48 scored snapshots needed).
Validation scoreboard
Model est-1.0+hl-2026-09-27 · realised Bybit liquidations · computed 2026-09-27T18:34:12+00:00| Horizon | Rank corr. (model) | Rank corr. (baseline) | Skill | Model ahead | Calibration | Log-lik. gain | Top-3 hit (model / baseline / random) | Scored snapshots | Coins | Verdict |
|---|---|---|---|---|---|---|---|---|---|---|
| 4h | — | — | — | — | — | — | — / — / — | 0 | 0 | insufficient data |
| 24h | — | — | — | — | — | — | — / — / — | 0 | 0 | insufficient data |
A verdict needs at least 48 scored snapshots per horizon. Skill above 0 means the model ranks the levels where liquidations happened better than the baseline; below 0 means it does worse. We publish the numbers either way.
How the estimate is built
Model est-1.0+hl-2026-09-27- Inputs. Open interest, price bars and volume that this site records from exchanges' public APIs, in 5-minute to 1-hour buckets.
- Opens. Every rise in open interest is booked as a new long and a new short (each contract has both sides) at the bucket's volume-weighted price.
- Liquidation levels. Each new position's liquidation price is spread over the distance-from-entry distribution below: longs below the entry, shorts above.
- Closes and decay. Falls in open interest close positions proportionally at every level, and all positions decay with a half-life, because most positions are closed without ever being liquidated.
- Clearing. When price trades through a level, the positions there count as liquidated and the level is removed.
| Setting | Value | Why |
|---|---|---|
| Model version | est-1.0+hl-2026-09-27 | stamped on every estimated chart; a new version is scored from scratch |
| Position half-life | 5 days | share of modelled positions still open after this long: half |
| Price bin width | 0.25% | log-spaced price levels |
| Maximum leverage | 100x | no liquidation level closer to entry than 1/leverage - margin |
| Maintenance margin | 0.5% | liquidation price = bankruptcy price adjusted by this rate |
| Open cap per bucket | 25% of OI | larger jumps (new listings, unit changes) are capped |
| Close cap per bucket | 50% of OI | — |
Where liquidation prices sit
Measured on real Hyperliquid positions| Liquidation distance from entry | Share of new positions |
|---|---|
| 0% – 2% | 12.0% |
| 2% – 5% | 7.0% |
| 5% – 10% | 20.0% |
| 10% – 20% | 18.0% |
| 20% – 50% | 17.0% |
The other 26.0% of new positions sit more than 50% from entry or can't be liquidated, and aren't placed on the map. Hyperliquid positions are on-chain, so their real liquidation prices can be measured; the same distribution is used for every coin and exchange. See the Hyperliquid liquidation map for real positions.
How it is scored
Daily, on Bybit data only- Ground truth. Bybit publishes every liquidation on its public stream (see data coverage), so the model is replayed on Bybit data alone and scored against Bybit's real liquidations.
- Snapshots. The model's state at the end of every UTC hour, from a 7-day replay, is compared with what happened over the next 4 and 24 hours. A snapshot is skipped if under 80% of that period has price data.
- Only levels price reached count. Levels are 0.50% wide; a level price never crossed can't show whether it was right, so it is left out. At least 5 crossed levels are needed for a rank correlation.
- Baseline. The same open interest spread over the same distance distribution around the current price, with no history. If the model can't beat this, its complexity isn't earning anything.
- Snapshot
- The model state at the end of each whole UTC hour, replayed on Bybit data only.
- Traversed bins
- 0.5% log price bins that price crossed within the horizon: long levels from the horizon's low to the price at the snapshot, short levels from that price to the high.
- Realised
- Bybit allLiquidation notional in the horizon, binned at the liquidation price (bankruptcy price adjusted by the maintenance margin rate).
- Baseline
- Open interest at the snapshot spread over the same distance prior around the current price, with no open-interest history.
- Rank correlation (rho)
- Spearman rank correlation between predicted and realised USD across traversed bins, per snapshot; the scoreboard shows medians.
- Skill
- Median of model rho minus baseline rho over the same snapshots; > 0 means the model beats the baseline.
- Calibration
- Sum of realised / sum of predicted USD over traversed bins (below 1: the map overstates, as vendors say theirs do).
- Log-likelihood gain
- Median log-likelihood gain of the model over the baseline, nats per realised USD.
- Top-k hit rate
- Share of the top-3 predicted bins per snapshot with at least $10,000 realised; the random rate is the share of all traversed bins.
Known limits
- Price bars are last-trade prices, not mark prices, which exchanges use to trigger liquidations.
- Opens and closes inside one open-interest snapshot cancel out, so fast in-and-out positions are invisible.
- One distance distribution is used for every coin and both sides.
- The scoreboard uses Bybit only; other exchanges publish a sample of their liquidations and can't be scored fairly.
- USD values on the heatmaps are modelled amounts; read them as relative intensity. Estimates may be materially wrong. Not financial advice.
See the estimates: liquidation heatmaps for every major coin.