How DeFi Liquidations Actually Work

Over-collateralised lending is the largest category in decentralised finance and the one whose failure mode is most consistently misunderstood. The mechanism is simple. The conditions under which it stops working are not.

The basic loop

You deposit collateral and borrow against it, always less than the collateral is worth. The protocol tracks a health measure — the ratio of your collateral value to your debt, adjusted by a factor reflecting how risky the protocol considers that collateral. When the measure crosses a threshold, your position becomes liquidatable.

At that point a third party, usually an automated bot known as a keeper, repays part of your debt and receives some of your collateral in exchange, plus a bonus. The bonus is the entire incentive structure: it is what makes someone spend gas racing to close your position.

Why the bonus size is a design decision

Too small and keepers do not act, because the profit does not cover gas plus the risk of losing the race and paying for nothing. Too large and borrowers are punished excessively for brief volatility, which drives them to competitors.

The critical property is that the bonus must exceed keeper costs under stress, not under normal conditions. Network congestion during a violent market move is exactly when gas is most expensive and when liquidations matter most. A bonus calibrated in calm conditions can become unprofitable precisely when the protocol depends on it.

How bad debt happens

Bad debt arises when a position’s collateral becomes worth less than its debt before anyone liquidates it. The protocol is then insolvent by that amount, and someone — an insurance module, a treasury, or ultimately depositors — absorbs it.

Three routes get you there.

Price moves faster than the machinery. Collateral falls far enough between oracle updates that the position goes from healthy to underwater without ever passing through a profitably liquidatable state.

The collateral cannot be sold. This is the most common real-world cause. A keeper liquidates by acquiring collateral they must then offload. If on-chain liquidity for that asset is thin, selling it moves the price against them and the trade is unprofitable — so they do not do it. The position is liquidatable in theory and untouched in practice.

Congestion. Liquidation transactions cannot land, or land too late, because the network is saturated by everyone else reacting to the same event.

The parameter that matters most

Attention usually goes to the collateral ratio. The more informative parameter is generally the oracle: how often it updates, how large a move it requires before updating, and what happens if it stalls.

A generous collateral ratio with a slow or infrequently-updating oracle can be more dangerous than a tight ratio with a responsive one, because the buffer exists to absorb the price movement that occurs between updates. Widening the buffer while widening the update interval can leave you no better off.

Reading a lending market’s risk

Useful questions, roughly in order of how much they tell you:

• What is accepted as collateral, and what is on-chain liquidity for it during a bad week rather than a good one?
• How concentrated are the largest positions? A handful of very large borrowers is a different risk profile from many small ones, because large positions cannot be liquidated without moving the market.
• Where do prices come from, how often, and what is the stall behaviour?
• Does the protocol support partial liquidation, or must a keeper close the whole position at once? Full-close requirements make large positions much harder to liquidate.
• Is there a mechanism for absorbing bad debt, and is it funded?

The general lesson

Liquidation systems are not primarily cryptographic mechanisms; they are market mechanisms that depend on someone finding it profitable to act quickly. Every failure of one traces back to a moment when acting stopped being profitable — and the design question is always whether the incentive survives the conditions it exists for.

This is not advice about whether to use any protocol. It is a description of the machinery so you can ask better questions about the ones you already use.