Every liquidation heatmap starts with the same problem: you know that open interest exists at a given symbol, but you do not know the leverage distribution behind it. A $500M open interest figure on BTCUSDT could be mostly 2x positions (low liquidation risk) or mostly 50x positions (high liquidation risk). The distribution determines where the forced moves are.
Most heatmap implementations solve this with a static assumption. They apply a fixed distribution curve, place estimated liquidation levels on a chart, and leave it at that. The problem is that leverage behavior varies by exchange, by market regime, and over time. A fixed model drifts from reality.
The three learning systems
The Heatmap Engine addresses this by learning from real events. Three feedback systems run continuously, each correcting a different dimension of the model.
Leverage Learner
When a liquidation event occurs on an exchange, the engine compares it against where the model expected liquidations to cluster. If the model expected peak liquidation density at $96,000 but the actual liquidation occurred at $96,400, the leverage distribution assumptions for that exchange and leverage range get adjusted.
Over thousands of observed liquidation events, the model converges toward the true leverage distribution for each exchange. Binance traders use different leverage profiles than Hyperliquid traders. The Leverage Learner captures these differences automatically.
Cascade Learner
Cascade chains follow patterns. When liquidations at one price level trigger liquidations at the next, the relationship between those levels contains learnable information: how much sell pressure from level one is required to push price to level two, what percentage of the exposed leverage at level two actually liquidates, and how quickly the chain propagates.
The Cascade Learner tracks these patterns across historical events and uses them to weight the probability of chain propagation at each level. It models chains up to 10 levels deep, with decreasing confidence at each subsequent level.
Exchange Calibrator
Every exchange runs a different liquidation engine. Margin requirements, liquidation formulas, insurance fund mechanics, and position size limits all vary. A single set of parameters applied across all 22 exchanges produces distorted results.
The Exchange Calibrator maintains per-exchange correction factors that account for these differences. It adjusts based on observed liquidation behavior on each exchange, ensuring that the heatmap for Deribit reflects Deribit-specific mechanics and the map for Bitget reflects Bitget.
Why static models drift
Market conditions change. During a low-volatility ranging period, traders use higher leverage because stops are tighter. During a trending market, leverage profiles shift. After a major liquidation event, surviving positions have different characteristics than the positions that were liquidated.
Continuous calibration
A static model treats last week's leverage distribution as today's reality. The Heatmap Engine updates with every observed event. The map you see at 14:00 reflects the liquidation events from 13:59 and all prior observations, not a fixed assumption from the initial deployment.
From assumption to observation
The core shift is from assumed leverage distribution to observed leverage distribution. Instead of guessing where liquidations will cluster, the engine measures where they actually occur and calibrates the map accordingly. The model improves with every real event.
This is why Glass is on version 14. Each version incorporated a new learning dimension or refined the feedback loop. The current engine processes data from 22 exchanges at $50 price granularity from 2x to 125x leverage, with continuous calibration from three independent learning systems.