How the stress numbers behind a collateral weight are measured from price history. Change an input and each formula, chart and weight below recomputes from the stored candle statistics.
A liquidation leaves the liquidator holding the seized collateral, and it has to sell it. How much it can sell before the price runs away depends on the bids resting in the order book, and those thin out exactly when the price is falling fast. For each asset this tab takes its worst independent 60-minute falls, the same falls that set the stress gap G, and reads the bid side of the book around them: how many dollars could be sold within 1% to 5% of the mid price in the 24 hours before the fall (normal), and at the thinnest moment inside it.
The fee is the same f as in the other tabs. It is what the liquidator earns on the sale, so a sale whose average slippage is above f loses money before gas, and a liquidator can be expected to stay away from it.
The depth archive is stuck for most assets on 9 to 11 October 2025, so the tables above cannot cover the largest crash. The executed-trade archive is intact, and it shows what real market sells paid. Each cell is the slippage of sell orders of that size in the 60 minutes from the start of the asset's worst fall that evening, against the prior 24 hours: the median, then the 90th percentile and the number of orders below it. A sell is the taker fills at one millisecond, so one order or a few, and its slippage is its average fill price against its first fill.
Where the book archive is intact that day (ZEC, LTC, UNI, QNT, PAXG) the book at its thinnest was 0 to 3% of normal. Realized slippage rose in the same direction but by less, because it only counts sells that were sent and filled, and it averages the hour instead of taking the worst minute. Read it as what a seller actually got, not as the worst the book could have done. Greyed cells have fewer than five orders. Red means the 90th percentile is above f.
Where the depth comes from. Binance's public archive of its USD-M perpetual order books: a snapshot about every 30 seconds with the total bid notional resting within 1, 2, 3, 4 and 5% of the mid price. That is the perpetual's book, not the spot book the candles come from, and for FARTCOIN, PUMP, SPX, XAUT, HYPE, MORPHO and ONDO the candles are Bybit's. Volume is from the candle source. Thinnest is the lowest 1% reading that at least two of any three consecutive snapshots (about a minute) reached, so one bad snapshot cannot set it. A shorter vacuum can still be missed, so it is not a floor.
Reading the numbers. Left is the thinnest book within 1% inside the fall divided by the normal book within 1%. Slippage walks the book from the mid price down, treating the liquidity between two levels as evenly spread, and gives the average price a sale gets, as a fall from mid; > 5% means the sale is bigger than the whole book within 5%. Max sale is the largest sale whose average slippage is still no more than f. Volume × is what traded in the 60-minute window, divided by the median 60-minute volume of the prior day. Median is over the asset's events and worst is the worst of them. The events are the worst falls in USD price, whichever debt asset is selected.
Realized slippage. From Binance USD-M aggTrades, the executed trades. Fills at one millisecond are one sweep, which can merge several orders and so overstate a sweep a little. The slippage leaves out the half spread. The sizes are the dollar value of the sweep.
Stuck feed. In some stretches the archive's 1% reading sits at one fixed value, or flips between two, for hours, even an hour before the move (BTC reads $2.14M against a normal $140M). That is a broken feed and not a thin book, so any stretch of eight consecutive snapshots with at most two distinct 1% readings is dropped, and an event needs most of its snapshots left. This removes 9 to 11 October 2025 for most assets, so the 10 October crash itself is missing for most of them, and the eight events shown are the worst falls that remain. Only falls inside the archive's history count (XAUT from March 2026, PUMP from its launch).
These are the same inputs as the lab, and they stay in step with it: change one here and the lab uses it too.
Click a column heading to sort, or an asset to open it in the lab. A red buffer or room means that check fails at these settings.
Picture a borrower who has borrowed the maximum against one collateral asset, and then the collateral price falls. Three price levels matter, and the two weights set where they sit.
Debt asset. With a stablecoin debt (aUSD) the debt price is flat, so only the collateral's price moves. With aBTC debt both legs move: a fall in the collateral ÷ debt price ratio is the collateral dropping or BTC rising. Every number on this page (σday, G, tw_lag) is then measured on that ratio, so a collateral that moves with BTC needs less buffer than the same collateral against aUSD.
Short version: room is for the borrower, buffer is for the protocol. Room is spent before the trigger, buffer after it. Raising the maint weight shrinks the buffer; raising init relative to maint shrinks the room. The clear line is a modelling threshold, not a protocol state: the program's own bankruptcy check only fires when what is left is close to zero.
Every one-minute close gives a log return. The standard deviation of those returns, scaled by the square root of the 1,440 minutes in a day, is the daily volatility. It sets the size of a normal day and the cushion a borrower needs.
Assume nobody can liquidate for a window W: the oracle is stale or the bot is down. For every minute i, measure how far the price falls from that close to the lowest low inside the window. Gemp is a high quantile of those falls. Choosing max uses the single worst fall in the whole history, the most conservative setting, which moves with every new crash.
If returns were normal, the gap at quantile q would be zq · σday · √(W/1440). The gap is measured to the lowest low in the window, so the relevant z is the two-sided one, zq = Φ−1((1+q)/2): a falling price reaches a given depth at some point in the window about twice as often as it ends the window there. Real prices fall further than that in the tail. The extra is the jump. The solver uses G = normal + jump, which is the larger of the two.
Initial health reads a smoothed price, maintenance health reads the live one. When the live price drops under the smoothed one, a position that looked fine at borrow time is already closer to liquidation. tw_lag is a quantile of that shortfall.
The buffer B is the price fall between the liquidation trigger and the line where a full liquidation stops clearing the debt. It has to cover G. The room R is the fall between the max-borrow price and the trigger. It has to cover R*.
Same ladder in money. Take a borrower at the max-borrow point with 100 of debt. The collateral is then worth 100 · wL / winit. As the collateral price falls, its value falls with it while the debt stays at 100 (debt price held flat, interest ignored). A liquidation only gets the collateral after the liquidator and insurance fees, so what it can actually repay is k times the collateral value. Three markers: the trigger is where the position becomes liquidatable; the clear line is where k × collateral drops below the debt, so a full liquidation no longer repays it; collateral = debt is where the collateral is worth less than the debt even with no fees at all.
Weights are the largest values on a 0.01 grid that meet both conditions. This page uses the same arithmetic as scripts/weight_calibration.py, applied to statistics computed once from the cached candles. Defaults for the fees are placeholders. Use the real bank values.
The 60-minute EMA span is a guess at the oracle's horizon, so check it against the feed. Candles come from one exchange and may differ from the oracle's own price. HYPE, MORPHO and ONDO use Bybit perpetuals, which track spot closely but are not spot; the others use Binance spot. Statistics for shorter histories are computed over that slice only.