If you read trade press on quant funds long enough, two names keep showing up under the heading "we know roughly what they do, we don't know exactly how." Quadrature Capital, in London. Radix Trading, in Chicago. Both pay the kind of headline numbers that make people read carefully, and both publish almost nothing.
This post pulls together what is actually public about both firms, then translates the techniques back down into things a person with a laptop and a Coinbase Advanced account can reasonably attempt. The honest answer is that you can copy the discipline, the data hygiene, and a few specific signal patterns. You cannot copy the latency, the capital, or the team.
What Quadrature Capital actually is
Quadrature was founded in 2008 by five employees who left G-Research, with trading commencing in 2010. The firm describes itself plainly: "a tech business" that "takes market neutral trading positions determined by statistical models" on its own proprietary capital. (Quadrature about page)
The firm is hyper-secretive. eFinancialCareers, which covers London quant pay, calls Quadrature "a bit of a cult" and notes it averaged £2.8 million per head across its 173 employees. The same outlet flagged a 2024 hire of Salvatore Scellato as head of research technology, joining from Google DeepMind, and the earlier hire of DeepMind alum Laurent Mazare in 2023 (Mazare left after four months to found an AI startup). (efinancialcareers Quadrature profile, efinancialcareers Quadrature DeepMind hire)
Public job listings and the firm's own site let you reconstruct the rough shape of the stack. Stated expertise covers statistical arbitrage, equities, technology, machine learning, AI, algorithmic and systematic trading. Reporting on infrastructure, again from eFinancialCareers, puts internal compute at over 20,000 CPU cores, 2,000 GPUs, 500TB of RAM, and 10PB of storage. The firm's published materials describe building "modular components shared across as many of their trading strategies as possible," which is the cleanest one-sentence summary of how a modern stat-arb shop scales: a single feature library, many strategy heads. (Quadrature home)
What you can fairly conclude from public material:
- Statistical arbitrage is the load-bearing strategy class.
- The team treats research compute as a first-class input, not a cost center.
- Hiring tilts toward research scientists who can engineer, with several DeepMind crossovers.
- The firm runs proprietary capital, not external client money, so it has no quarterly investor narrative to manage.
What Radix Trading actually is
Radix is American, headquartered in Chicago, with offices in New York and Amsterdam. It was co-founded by Benjamin Blander, who previously ran high-frequency trading at Citadel, and Michael Rauchman, formerly chief technology officer at GETCO. Radix describes itself as "a research firm, powered by technology and monetized through trading." (FIA new-member profile of Radix)
Two phrases recur across Radix's public surface: "innovative machine learning and statistical methodologies" and "research-driven, algorithmic trading strategies across major electronic markets worldwide." The firm pitches itself as a place where the round-trip from idea to live execution is short, with continuous enhancement of its automated research platform. (Radix Trading site)
Hiring is the most useful tell here. Quantitative researcher postings on math-job boards confirm that Radix recruits PhDs, postdocs, and professors from mathematics, physics, statistics, and computer science, then pairs them with engineers. Blander himself holds a PhD in mathematics from the University of Chicago. (Radix QR posting on MathJobs)
What you can fairly conclude:
- Strategy origin is academic-style research, not chart-pattern intuition.
- Coverage spans major electronic markets, not a single asset class. That implies a generalized infrastructure layer that abstracts venues from strategies.
- ML is in the loop, but described alongside "statistical methodologies," which usually means classical econometrics is still doing real work. The firm is not selling a deep-learning brand.
The shared shape underneath
Strip the marketing language off both firms and the shape of the work looks similar.
Each firm runs its own capital. That removes the constraint of explaining a drawdown to an LP and lets the research process target raw expected value rather than smoothed reported returns.
Each firm treats infrastructure as part of the strategy. Quadrature's published compute footprint and Radix's "automated research platform" language both describe environments where a researcher can iterate fast and a winning iteration can be deployed without a six-week handover. That iteration loop is the actual edge in many shops, more than any single signal.
Each firm hires for the ability to do science, not for the ability to talk about markets. Quadrature is pulling DeepMind people. Radix advertises explicitly to PhDs, postdocs, and professors. Both then expect those hires to write production code.
If you wanted to compress all of this into one rule, it would be: the edge is in the research process, not the trade. The trade is the output of a system you keep refining.
What a retail investor can reasonably borrow
You are not going to colocate at NY4. You are not going to buy 2,000 GPUs. The list below is what you can actually replicate from a kitchen table.
1. One feature library, many strategy heads. This is the single most portable idea. Define your universe once. Compute a small set of features once: volume z-score, return z-score over multiple horizons, realized volatility, range expansion, simple regime flags. Then express every strategy as a thin function of those features. You will move faster, you will not double-implement bugs, and you can backtest a new strategy in an afternoon instead of a week.
2. Statistical-arbitrage thinking, even on one asset. Stat arb at Quadrature scale is pairs and baskets across thousands of equities. At retail scale, the useful inheritance is the idea of mean reversion around a model. Build a simple expected-price or expected-spread model, trade the residual when it gets unusually wide, do nothing when it doesn't. The point is not the trade, the point is having a clear definition of "wide."
3. Research compute as a first-class input. A consumer GPU and a properly indexed Parquet store will run rings around a thousand-line spreadsheet. If you are doing this seriously, give yourself a real research environment. JupyterLab, a folder of cleaned data, a notebook per strategy idea, and an automated daily refresh pipeline.
4. Walk-forward validation, not in-sample story-telling. Both firms are research shops, and a research shop survives by not lying to itself. Out-of-sample backtests, walk-forward windows, transaction-cost-aware Sharpe, and explicit regime breakdowns. If your backtest looks great over the whole window but falls apart inside any individual year, the whole window is the lie.
5. Cost-aware trade decisions. Cite this and then act on it: Han, Kang, and Ryu (2023) found that many published cryptocurrency momentum strategies stop working once realistic transaction costs are included. The same pattern shows up in equity stat-arb papers. If your edge is 12 basis points and your cost is 8, you have 4 basis points of expected return and an enormous variance burden. Most retail traders do not subtract their costs honestly. (Han, Kang, Ryu, SSRN 4675565)
6. Sharp-trader tracking on prediction markets. This is the single retail edge that has actually grown in the last two years. The top sliver of Polymarket wallets, when their behavior is tracked carefully, leak useful information minutes to hours before public news catches up. The framework I use for this is the Whale Intelligence Polymarket integration approach: monitor on-chain Polymarket flow, define a "sharp" wallet by lifetime PnL plus consistency, and treat clusters of three or more sharps in the same direction within 24 hours as a tradeable signal. None of the elite firms are doing this. It is genuinely retail-shaped work.
What you cannot borrow
Latency. If a strategy decays in milliseconds, you do not have it. Quadrature, Radix, Citadel Securities, Jane Street, and HRT compete on the order of microseconds. There is nothing in the retail toolchain that closes that gap, and there is no reason to try.
Capital base. A 4 percent expected return on $5 billion is $200 million. The same 4 percent on $50,000 is $2,000. Many strategies that work for institutions are uneconomic for retail because the absolute dollar return cannot pay the fixed cost of running them.
Regulatory access. Direct market access, primary dealer status, ETF authorized-participant rights, and CFTC futures-commission-merchant relationships are not retail products. They give institutional players legal pathways to liquidity that retail investors simply cannot reach.
Team. This is the one most people underrate. The version of Quadrature you read about, with £2.8 million per head and 173 people, has hundreds of researcher-years of tacit knowledge baked into how they pick problems. You can replicate the process. You cannot replicate the institutional memory.
A practical retail program for the next 90 days
If you read the above and want a concrete set of moves, this is what I would actually do:
- Pick one venue and one universe. Coinbase Advanced top 20 USD pairs is a fine starting set, because data is free, fees are public, and the API is documented. (Coinbase Advanced API Python SDK)
- Build the feature library before you build any strategy. Returns at 1, 6, 24, and 168 hours. Volume z-scores. Realized volatility at the same horizons. A simple regime flag based on Bitcoin's 30-day return.
- Test a single time-series momentum strategy with realistic fees. Not because momentum is going to make you rich, but because if you cannot get an honest backtest of momentum to match published results, you have a measurement problem and you need to fix that first.
- Add Polymarket sharp-wallet monitoring as a parallel research thread. Different data, different infrastructure, different decay profile. Run it as a notebook for the first month before you put any capital on it.
- Track every trade against your model's expectation. Not P&L. Expected versus realized. That is the only feedback loop that improves the system.
Three months of that beats six months of switching frameworks. The two posts that follow this one cover (a) the broader landscape of why elite desks consistently outperform, and (b) a specific GitHub stack you can actually clone and run, with a working backtest example.
Related reading
- Why Top Quant Desks Outperform And How U.S. Tax Code Multiplies The Edge for the structural side, including Section 1256 and the Lender Management family-office ruling
- The Retail Quant Stack: Coinbase, ccxt, vectorbt, And A Working Backtest for the runnable code companion to this piece
- What Actually Fixed My Claude Code Sessions for the research-discipline habits that translate to trading work
- A 24.8 Percent AI Crypto Portfolio, Back-Checked for the corresponding skeptic's checklist when reading return claims
Fact-check notes and sources
- Quadrature Capital company background, infrastructure, and per-head pay: efinancialcareers profile, efinancialcareers DeepMind hires, Quadrature about page, Quadrature home
- UK Companies House record: Quadrature Capital Limited filings
- Radix Trading background: FIA new-member profile, Radix Trading site, Radix Wikipedia entry, MathJobs quantitative researcher posting
- Crypto momentum after realistic costs: Han, Chulwoo; Kang, Byeongguk; Ryu, Jehyeon (2023). Time-Series and Cross-Sectional Momentum in the Cryptocurrency Market: A Comprehensive Analysis under Realistic Assumptions. SSRN 4675565
- Coinbase Advanced API: coinbase/coinbase-advanced-py on GitHub
- Whale Intelligence framework reference: source materials in the author's personal research archive, summarized in this site's Claude trading lessons post
This post is informational, not investment advice. Quadrature Capital and Radix Trading are private firms with limited public disclosure; details are derived from public reporting and corporate filings as cited. Past performance of any cited strategy or firm is not indicative of future results. Trading securities, futures, derivatives, and cryptocurrencies involves risk of loss. Consult a qualified financial advisor before allocating capital to any active strategy.
