Real Profitable HFT Market Making Lives or Dies by Infrastructure

You backtested your shiny Avellaneda-Stoikov (or Guéant-Lehalle-Fernandez, or whatever flavour) model and got a Sharpe of 15. Congrats!

Now try running it live with 500 μs latency while Jane Street is operating at 87 nanoseconds in the same book.

You’re not competing. You’re free alpha for everyone faster than you.

In high-frequency market making, the model is maybe 10–20 % of the battle. The other 80–90 % is pure infrastructure and devops on steroids.

Here’s the no-BS breakdown of what actually matters in 2025.

> FIG: Market Making: Avellaneda–Stoikov model

The Latency Hierarchy of Pain

If you’re not in the first two rows on liquid instruments… just don’t.

The Real Stack – What Winners Actually Use

Real 2025 example stack for a profitable independent shop on Nasdaq/ CME:

> FIG: Tick-to-trade

Total tick-to-trade: 80–150 nanoseconds on a good day.

Language Choice – Where Python Dies

• Python → research, backtesting, crypto toys only
• C++ / Rust → production quoting engine (if not on FPGA)
• Verilog / SystemVerilog → the real winners write the entire strategy in hardware

A single garbage collection pause of 200 μs just wiped out your entire day’s P&L.

The Hidden Killer: Queue Position & Adverse Selection

Even with perfect latency, if you don’t track your exact position in the price-time priority queue, your model is lying to you.

Example:

• You think you’re first in line at the bid → quote aggressively
• Actually you’re 50th → every informed seller hits you first → you get run over

Top shops track every add/cancel on the wire and maintain their own shadow book with sub-microsecond accuracy.

Without this, AS (or any model) overestimates profits by 5–20× in real markets.

Crypto vs Traditional – Slightly Different Rules

Crypto is more forgiving because:

• 24/7 markets (no end-of-day inventory panic)
• Higher volatility → wider spreads → more room for latency slop
• Many venues are still slow (Binance spot can be profitable with 5–20 ms)

But the big boys (Wintermute, Jump Crypto, Cumberland) are already running the exact same FPGA/microwave/colocation game on centralized exchanges and on-chain (MEV, Solana Jito bundles, etc.).

Bottom Line – Can a Solo Dev Win?

Yes… but only in niches:

• Illiquid altcoins
• Emerging L2s / new perp DEXs
• Geographic arbitrage (e.g., Korea premium)
• On-chain market making where latency is measured in block times

On BTC/USD or ES futures? Forget it unless you have $5M–$50M war chest for infrastructure and a team of ex-Citadel FPGA wizards.

Final Reality Check

Beautiful model + mediocre infrastructure = consistent losses
Decent model + god-tier infrastructure = printing money

Choose wisely where you spend your next 12 months.

Next post: “How I built a sub-millisecond market maker in Rust for under $10k” (spoiler: it makes $3/day on some random altcoin… but it’s fun).

Stay fast, or stay home 🚀