Manifold’s early trading edge came from avoiding a direct fight with firms such as Jump, Tower, and Jane Street on centralized exchanges and moving into CEX-DEX arbitrage, where decentralized venues were still structurally inefficient, founder Jae Chung wrote in a first-person account published by TechFlow.
Chung said he started Manifold in 2021 at age 21 with no real background in quantitative trading. Before that, he had entered crypto in 2016/17 as a white-hat hacker, ran validator nodes for several years, and made a number of trades and investments during DeFi summer and the NFT boom. None of that, he wrote, had much to do with quant trading, and advisers around him told him the idea was a bad one because he did not understand the competitive field he was walking into.
Early attempts failed before the team found a workable direction
Chung said many people assume quant trading is simply a matter of finding a pattern and making money. His experience was the opposite. Manifold tested a wide range of ideas at the start, including arbitrage, spread capture, basis trading, statistical arbitrage, yield farming, and even deep learning. None of those approaches worked at first.
His conclusion was that trading is a zero-sum game. If too many people are better than you at the same game, there is very little left to take. That forced the firm to focus less on broad experimentation and more on choosing a market where it could build a real edge.
Why DeFi looked like the better table to sit at
Chung grouped the strengths of top trading firms into three buckets: speed and latency through trading infrastructure, proprietary data and order flow, and research plus alpha generation.
On centralized exchanges, he said, those strengths were already concentrated in firms such as Jump and Tower. They had state-of-the-art low-latency systems deployed across major exchanges including Binance, OKX, Bybit, and Coinbase. They also had fee tiers Manifold could not reach at the time because of lower trading volume, as well as client relationships that gave them order flow and data advantages.
Manifold, by contrast, had only entry-level trading and research infrastructure and no large army of engineers and researchers. That pushed the company to find a different game. Chung said the answer was DeFi, where most decentralized exchanges used AMM pricing rather than order books and where the trading stack had to handle pricing, liquidity ranges, slippage, transaction submission, execution, and confirmation on-chain.
He argued that top firms were not pushing as aggressively into that market then, possibly because it sat outside their established infrastructure, because the market was smaller than the classic CEX venue set, because DEX-related regulation was still uncertain, or because those firms were already making enough money elsewhere.
The talent profile also fit Manifold better. Chung described the team as a mix of crypto-native engineers who understood smart contracts and on-chain execution, working alongside quant researchers from more traditional backgrounds such as Citadel and Tower.
He gave one hiring example from later on. One of Manifold’s strongest employees had been a data engineer at an insurance company while also running small atomic arbitrage strategies as a solo MEV searcher on Polygon. Before he joined, Chung communicated with him through the anonymous email address 0xaddress@protonmail and reviewed the bot’s transaction history on a block explorer.
The first prototype was written in TypeScript
Before shifting the whole firm toward DEX trading, Chung said he built a basic on-chain trading system entirely in TypeScript to test whether arbitrage between centralized and decentralized exchanges could work. The setup compared Binance and FTX against the top ten blockchains by liquidity. Internally, the team called the strategy CEX-DEX arbitrage.
The opportunity existed because blockchains batch or delay transactions within block times, which could be around 200 milliseconds on some chains, 1 to 2 seconds on others, and several seconds elsewhere. That is much slower than quote-by-quote price updates on centralized exchanges, creating a lead-lag relationship. At the same time, much of the liquidity on decentralized exchanges was passive or stale and not actively managed by bots, while price discovery still happened mainly off-chain.
Even though the prototype was rough and not built around extreme latency sensitivity, Chung said it still captured real arbitrage opportunities and remained profitable after fees. After months of research with little progress, that was the first sign the approach had room to scale.
Once the prototype worked, execution became the main job
Chung said it is unusual for an arbitrage prototype to work at all, because arbitrage, by definition, is profit lying on the ground and usually somebody faster is ready to take it first. But once Manifold had a live setup that made money, the path forward was clearer.
The firm identified several upgrade paths:
- rewrite the TypeScript system in a faster language
- add more DEXs per chain and expand to more chains
- improve the math around position sizing and slippage
- optimize fees to raise profit per trade
- reach better fee tiers on centralized exchanges as volume increased
- cut gas costs through on-chain engineering
- improve execution quality and fill rates
- optimize inventory to improve capital efficiency, returns, and uptime
He said the appeal of high-frequency trading is the speed of market feedback. Once a change is implemented and put into production, the firm sees very quickly whether profitability improves or not. That created a fast loop between engineering work and dollar outcomes.
Chung also recalled a specific stretch of that period. On Dec. 24, 2021, he and co-founder Sid were coding overnight in their apartment office and did not realize Christmas had arrived. At around 1 a.m., Chung looked at the time and pointed it out. About 30 seconds later, Sid replied, 「Oh. Merry Christmas.」 They kept coding.
From a few million dollars deployed to more than $10,000 a day
After a series of broad improvements, Chung said the CEX-DEX arbitrage strategy began making more than $10,000 per day with only a few million dollars deployed.
The firm’s early priority was to make a new system, written in Go and Solidity, good enough to scale across multiple chains and to 10 to 20 DEXs on each chain. Manifold also integrated newer decentralized exchanges using the UniV3 tick pricing model. Those venues were more efficient because of concentrated liquidity and generally had lower slippage.
As scale increased, so did the firm’s volume on centralized exchanges, which helped lower trading fees. With more capital available, the strategy also expanded into more pools and more pairs.
Competition changed the game
As the easy opportunities were taken, Chung said competition rose quickly. On some chains, Manifold could no longer wait for a 10 basis point spread because other bots were willing to trade at tighter levels.
He offered a simple example. If ETH was trading at $2,000 on Binance and $2,001 on Quickswap on Polygon, the spread was 0.05%, or 5 basis points. A bot configured to trigger at that level, after accounting for fees and slippage, would sell ETH on Polygon and buy ETH on Binance, restoring approximate parity. A slower or more expensive bot waiting for a 10 basis point spread would never see the trade, because the gap would be closed before Polygon ETH reached $2,002.
The cost of executing each trade varied from one firm to another, he said, based on three main factors: CEX fee tiers, whether the firm was taking or making on the centralized exchange side, and gas costs on-chain.
Manifold also started running into firms such as Wintermute and had to engage in priority gas auctions, or PGAs, against sets of bot addresses. Chung described this as an MEV-derived technique where a trader repeatedly resubmits the same transaction with a higher priority gas fee to improve ordering within a given block.
To compete in that environment, he said, a firm needed to do three things well: estimate the profit of a specific trade with high accuracy, run infrastructure that could land a transaction within one block after spotting the opportunity, and know which addresses it was bidding against while quickly determining whether it had won or lost.
He added that similar methods appear across different chains depending on how transaction ordering and submission are handled, and that arbitrageurs must keep adapting as blockchains change rules meant to redirect value toward users, protocols, or the chains themselves.
Gas optimization opened trades others could not see
Chung said the total gas cost of an on-chain trade is roughly a function of gas used multiplied by gas price, with gas price driven by network congestion. Priority fees matter in PGA-style competition, but once a trader is already able to land a transaction within a block, there is only so much room left there.
The larger gains came from reducing gas usage itself. Lower gas consumption cuts costs linearly. One technique the firm used was efficient bit packing. Because the Ethereum Virtual Machine stores data in 32-byte, or 256-bit, slots, smaller data types can be placed next to each other so Solidity packs them into a single slot, reducing the cost of reads and writes such as SSTORE and SLOAD.
Chung said the result was meaningfully less zero padding in transactions, which reduced transaction size and total fees. That let Manifold take trades that were effectively invisible to competitors with higher gas costs. In his example, a trade that still produced a $1 net profit for Manifold after fees might produce $0 for a less optimized rival.
That also meant the firm could bid more aggressively to win an arbitrage and stay profitable. Over time, more wins led to more volume, and more volume in turn improved CEX fee tiers and allowed the firm to capture larger arbitrage opportunities per trade.
Capital efficiency became the next bottleneck
As the strategy improved, Manifold expanded across more chains, DEXs, and trading pairs. Chung wrote that putting most of the firm’s capital into CEX-DEX arbitrage made sense at the time because the strategy had an annualized return above 100% and no down days.
Still, the setup had a structural problem. Capital had to be spread across venues, and inventory had to be held in multiple pairs. Chung used ETH/USDT arbitrage between Binance and Arbitrum as an example. If the firm started with $500,000 of ETH and USDT on each venue and Binance traded at a premium, the bot would sell ETH on Binance and buy ETH on Arbitrum. If that premium persisted, the firm could run out of ETH on Binance and out of USDT on Arbitrum very quickly.
Centralized exchanges offered margin, which allowed some continued trading when inventory skewed too far. On-chain, though, the bot would stop once inventory was depleted. To keep trading, Manifold had to rebalance by moving excess USDT from Binance to Arbitrum and excess ETH from Arbitrum back to Binance.
Even with 24/7 rebalancing, the firm still missed some of the most profitable moments because arbitrage PnL clustered during high-volatility periods when spreads widened rapidly. In those conditions, $500,000 of inventory could be consumed in seconds. Chung said exchange withdrawal times, already around five minutes in normal conditions, became even slower during volatile periods.
The first answer was an automated system called Hydra, which attempted to move capital continuously by using blockchain bridges and CEX withdrawal APIs. It was supposed to detect surplus assets in some places and move them to venues that were short. Chung said that version did not work as planned. Bridges were unreliable, assets could be lost or disappear, and the team had to track manually which bridge had failed and when. Bridge transfers were also too slow during active markets.
Hydra v2 used Aave to solve inventory shortfalls on-chain
Hydra v2 moved away from bridges. Instead, Manifold integrated Aave, or a similar protocol where Aave was not available, to create an internal on-chain spot margin system.
Chung described the idea as simple. In the earlier Arbitrum example, repeated ETH buying would leave the firm with too much ETH and too little USDT. Hydra v2 would detect excess ETH, lend part of it on Aave, and borrow USDT against that collateral.
He said the implementation was conceptually straightforward, but the PnL impact at the time was one of the largest improvements the firm made. Looking back, Chung wrote that none of the pieces individually required something like rocket science, but it would have been unrealistic to invent them all from the start. The key was getting to a live strategy first, then iterating relentlessly by monitoring fill rates and profitability and letting market feedback point to the next constraint.
A profitable alpha for years, then a gradual fade
Chung said CEX-DEX arbitrage remained Manifold’s main profit engine for a long stretch. In its strongest period, days with large market swings could produce more than six figures in profit. The firm also built infrastructure generic enough to connect to a new chain within one hour, which often made it one of the earliest arbitrageurs on newly launched or newly relevant chains when market inefficiencies were highest.
But he stressed that, like almost every alpha, the edge decayed with time. Arbitrage in particular invites competition that compresses spreads until profit margins are much thinner. In the earlier PGA example, if two firms had identical gas optimization and the same CEX fee tier, they would eventually bid up to the lowest profit level each was willing to accept because their execution costs were the same.
By late 2025, Chung wrote, CEX-DEX arbitrage was no longer the unusually high-return strategy it once had been. Even so, he said the firm already had other stable, high-return strategies running in the background, which allowed it to redirect time and capital elsewhere. He added that CEX-DEX arbitrage still holds a special place for him because it gave Manifold its first legitimate foothold and a base from which to build its position.
The account was written by Manifold founder Jae Chung and published in Chinese translation by TechFlow.
