HFT Market-Making Systems
I spent the first 13 years of my career designing algorithmic trading systems for large HFTs in options, futures, and ETFs (intl/US equities, commodities, fixed income, volatility). Here’s a summary of each component, note the role AI/ML actually plays.
Market data. A market-making system requires realtime orderbook deltas and trades, not just for the symbol that’s being quoted but also for every related instrument used in computing the theoretical fair value of the instrument. Every exchange has a different protocol specification and format, from JSON over websocket (slowest) to FIX (common) to fixed-width binary (fastest). The lowest latency systems will ingest market data on a different thread or process to avoid interfering with the core trading logic of the system.
Fair Value. Successful market making depends on a precise understanding of the theoretic fair value of an instrument, which is the indifference point between buying and selling the instrument modulo fees and other costs. The complexity spans from a several-parameter model based on one or two instruments to highly multivariate calculations across a basket of different symbols. This component is usually the firm’s most guarded intellectual property.
Order placement. Accurately predicting the price of an instrument is only half of the core logic of a trading system; deciding where and how to place orders is equally determinative of P&L. Maintaining balanced two-sided quotes, optimizing for queue position, minimizing traffic to comply with order rate limits, canceling to avoid adverse selection, and order sizing are all intricately determined by the strategy designer.
Exchange connectivity. Similar to the market data component, all trading systems must send orders and cancels using each exchange’s unique messaging protocol. JSON, FIX, and binary formats are all used, with the last most common among traditional exchanges where latency and throughput are optimized. The connectivity layer can be embedded within the order placement process or segregated as its own gateway process.
Offline training. Training is not part of realtime transformation of ticks to trades, but it is one of most important differentiators between trading firms. All parameters that define fair value calculation and order placement are set via offline compute tasks that are often guided by machine learning. A significant amount of the GPUs that large HFTs buy or rent are directed to model discovery and parameter optimization.
A handful of HFTs already use specialized GPUs for the linear algebra behind valuation, but the practice remains rare, confined to strategies where predictive power matters more than shaving the last nanoseconds off latency. As purpose-built inference silicon such as Google’s TPU and OpenAI’s Jalapeño becomes widely available, more HFTs will fold online ML and inference into their systems’ trading loops.
How to maximize your cashback as a non-USD/EUR bro.
I use 3 cards, and they all have cashback caps:
1. Gnosis: 3% up to $350/week (0% FX)
2. EtherFi: 3% up to $2k/month (1% FX)
3. Plasma: 3% up to $1k/month (1.5% FX)
I top up $350 into the Gnosis Earn vault.
Once that’s exhausted, I switch to LiquidUSD on EtherFi.
Rinse and repeat every Monday (that’s when Gnosis resets).
If I happen to exceed $2k on EtherFi, Plasma enters the rotation.
Further expansion via the wife’s cards is also available 😅
If you however are a USD/EUR bro you can just use EtherFi. As long as you spend more than $3.5k/month, you qualify for Lux tier and get 3% cashback on up to $10k of spending.