Welcome to the next crypto level
← Nexorena

DeFi & stablecoins

Developer compares routing on 3,429 orders drawn from CoW Swap

Rohan Indurkar published a test of his Split Optimizer engine using sell orders based on CoW Swap trades. The results measure quotes after network costs under defined conditions. They represent neither realized gains nor an integration with the CoW protocol.

Nexorena Editorial Team

Developer compares routing on 3,429 orders drawn from CoW Swap

Quoting the same orders at the same blocks

On November 20, 2026, developer Rohan Indurkar presented a test of Split Optimizer on the CoW DAO forum. The engine splits an order across multiple pools or paths. The sample includes 3,429 sell orders. Asset pairs and sizes come from executed CoW Swap trades, with amounts ranging from about $13,000 to $12 million and a median of $72,000.

On November 18, three decentralized exchange aggregators quoted the orders live. According to the author, Split Optimizer then used the same chain state each aggregator saw for the corresponding order. The comparison covers only types of on-chain liquidity available to his engine. Requests for quotes and private market makers were excluded, and aggregator fees were added back to the quoted amounts to remove their effect from the test.

The average difference varies by aggregator

After network costs and weighting by order size, the author found a 4.8-basis-point advantage across 3,303 orders against the first aggregator and a 1-basis-point advantage across 3,289 orders against the second. The third was compared on just 656 orders of at least 100 ETH. The weighted advantage was 0.2 basis points across volume, even though the engine returned a worse quote on 395 orders. The largest orders explain the difference between those measures.

The author proposes running his engine alongside an existing router and selecting the better quote for each order after network costs. On the sample, that calculation produced advantages of 5.3, 1.8, and 1.1 basis points against the three respective aggregators. Selection used the test quotes. It does not measure how many of those quotes would have led to a trade.

A method still to be tested in production

The engine builds on Fynd and Tycho, two open-source components from PropellerHeads, and uses a proprietary optimization algorithm. The author reports a median computation time of around 120 milliseconds per quote on one virtual machine with 16 virtual CPUs. The measurement excludes the full time needed to execute an order and price changes between quote and settlement.

Indurkar published references to the transactions and blocks used, along with a hash of the results and collected raw responses. The aggregators are anonymized in the publication. The test offers a starting point for comparing the method against the same chain data. A production evaluation would still need to measure quote availability, end-to-end latency, and the amounts actually received.

Article sources

Written with love byRyan MitchellMeet this editorial pen name ↗

Explorer Nexorena