Blog
Blog
On a centralised exchange, price is discovered by traders posting bids and asks, then choosing when to be aggressive. On an automated market maker, price is set by a formula. No bids. No asks. No market makers in the traditional sense. The mechanics that govern where a trade executes - and what happens to price after it - are structurally different from anything in CeFi, and most quantitative frameworks treat the two environments as equivalent.
That equivalence is wrong, and it has a cost.
A centralised exchange (CeFi) matches buyers and sellers. Someone posts a limit order. Someone else crosses it. The match creates a trade. Nothing happens to prices until humans or algorithms choose to act.
An AMM operates differently. A liquidity pool holds two assets in fixed proportion, governed by an invariant curve. Price is not posted by anyone. Price is derived from the ratio of reserves at the moment a trade arrives. A buyer swaps token A for token B. The pool delivers B and receives A. The ratio shifts. The new price is a mathematical consequence of the new ratio. The next buyer pays more for B than the last one did.
Every AMM trade is a guaranteed loss for the trader who initiates it - the formula ensures the pool extracts value - and yet billions of dollars flow through them daily because the alternative, centralised custody, is the larger risk for many participants.
The practical implication is immediate: there is no order flow imbalance to measure, because there are no orders. There is no aggressive-buy versus aggressive-sell decomposition. There is only: a trade happened, it moved the ratio by this much, and the implied price is now here.
On CeFi, the bid-ask spread carries information. Cont, Kukanov, and Stoikov demonstrated in 2014 (Journal of Financial Econometrics, 12(1), 47-88) that order flow imbalance at the top of the book has statistically significant predictive power for mid-price changes over the next seconds to minutes. The insight is that the spread reveals intent before price moves.
On an AMM, there is no spread in that sense. The pool price and the reference market price are the same thing only when arbitrageurs have just corrected them. Between corrections, the pool can be arbitrarily stale. A centralised reference price might move five percent while an AMM pool's implied price does not change at all - because no one has submitted a trade to the pool.
The entity that keeps AMM prices current is the arbitrageur. When CeFi price runs ahead of the AMM pool, arbitrageurs buy the cheap token in the pool and sell it on CeFi. That trade corrects the AMM price and extracts value from liquidity providers in the process. This extraction is called impermanent loss in the protocol literature. It is not impermanent in any meaningful sense. It is a structural, recurring transfer from passive liquidity providers to arbitrageurs, every time a significant price move occurs on CeFi.
What this means for signal detection: AMM price changes are not continuous price discovery. They are discrete correction events, triggered externally, with a lag that varies based on the size of the price discrepancy and how competitive the arbitrage environment is. The AMM pool price is always a lagged, discretely updated reflection of CeFi price discovery. The signal is not in the AMM. The signal is in the CeFi orderbook that drives the corrections.
The absence of traditional order flow does not mean AMMs carry no microstructure information. It means the relevant signals are different.
Pool reserve ratios before and after large trades reveal directional pressure that has not yet been corrected by arbitrage. When a large swap moves an AMM pool significantly, the implied price gap between the pool and CeFi reference price becomes a measurable signal: arbitrageurs will close that gap, and how quickly they do reveals something about available capital and attention in the arbitrage environment.
Gas price and transaction queue depth on networks like Ethereum carry information that has no CeFi equivalent. A spike in pending transactions targeting a specific AMM pool during a major price move signals that many arbitrageurs identified the same opportunity simultaneously. The arbitrageur who wins is the one who paid highest priority fee. The losers' transactions fail or land at worse prices. This queue is publicly observable before transactions confirm, which creates a signal window with no analog in any centralised exchange.
Block timing matters in a way it simply does not in CeFi. AMM trades execute in discrete blocks. Everything that happened to the pool during one block resolves simultaneously at block confirmation. The maximum extractable value literature documents how block proposers can reorder, insert, or censor transactions to capture arbitrage profits - a practice with no equivalent on any centralised exchange. Strategies that treat on-chain timestamps as equivalent to CeFi event timestamps will misattribute causality, because the block structure means simultaneity on-chain and causality off-chain are different concepts.
Most DeFi tokens also trade on CeFi, or have synthetic equivalents that do. The price relationship between the two venues is the highest-signal observable in the cross-market structure.
When CeFi price for a DeFi token runs ahead of on-chain pool prices, arbitrage flow is about to hit the AMM. The AMM price will correct. Liquidity providers will absorb losses. That sequence is predictable from the CeFi signal before it resolves on-chain. Strategies watching only CeFi can see it forming. Strategies watching only the AMM see it only after it has already happened.
The reverse is more interesting and less observed. On-chain trading activity in a protocol - governance token accumulation, new liquidity pool deposits, whale-wallet movements - sometimes precedes CeFi price moves. The on-chain data is public, but parsing it in real time requires indexing full chain state, which is an infrastructure problem several orders of magnitude harder than connecting to a CeFi WebSocket feed. Most market participants do not solve it. The information asymmetry between those who do and those who do not is structural and does not compete away easily.
Large on-chain transactions that move an AMM pool price significantly are immediately visible in the mempool before they confirm. This is not the case on any centralised exchange. Order flow on CeFi is opaque until the trade hits the tape. On-chain order flow is transparent before execution. The engineering is non-trivial, but the information advantage is real.
AMM microstructure is not a better edge than CeFi microstructure. It is a different edge, with higher infrastructure requirements, lower liquidity at comparable scale, and a more complex causality structure where on-chain and off-chain signals interact in ways that any single-venue framework will misread.
Treating AMM price data as if it were a CeFi orderbook - running OFI calculations on swap events, looking for aggressive-buy and aggressive-sell decompositions, applying spread-based metrics - produces numbers that look valid and mean something different from what the analyst assumes. The mechanism that generates the data is different enough that the interpretation must be rebuilt from the mechanism, not borrowed from CeFi intuition.
Most quant frameworks were built for CeFi and extended to DeFi by substitution rather than redesign. The substitution works until it encounters the structural cases where AMM mechanics diverge: corrective arbitrage lags, block-level simultaneity, mempool transparency, and the absence of anything resembling a limit orderbook. At those moments, the CeFi-derived signal gives the wrong answer.
Price discovery on centralised exchanges is a conversation between liquidity providers and aggressors, running continuously, measurable in real time. Price discovery on AMMs is a formula that gets corrected by arbitrageurs, whose timing and aggressiveness are driven by what is happening in CeFi.
The relationship runs one direction cleanly: CeFi discovers price, AMMs follow. The interesting exceptions - when on-chain activity precedes CeFi moves - are real and measurable but require chain-level infrastructure that most participants never build. The more common mistake is running CeFi microstructure analysis on AMM data and finding signals that are artefacts of arbitrage correction patterns rather than genuine market pressure. That is also why the crypto orderbook fragmentation problem remains relevant: even on the CeFi side, "the market" is already split across venues before the DeFi leg enters the picture. If the CeFi reference feed is too stale, why 100ms is an eternity in orderbook data explains why the arbitrage window you thought existed may already be gone. And for the part of the auction CeFi still does own, what market depth actually measures defines the depth context the AMM is reacting to rather than creating.
Two environments. One underlying asset, often. Completely different mechanics. The signals that work in one environment will not, in general, work in the other.
Because price is derived from reserve ratios and pool formulas, not from participants posting and cancelling bids and asks.
Usually on CeFi orderbooks, with AMMs catching up through arbitrage when the pool price diverges from the broader market.
Because on-chain flow can be seen before confirmation, which creates timing and ordering dynamics with no direct CeFi equivalent.
Not safely. The mechanics are different enough that the same labels can point to different causes.
Treating AMM swap events as if they were equivalent to aggressive orderbook flow instead of understanding them as formula-driven correction events.