In this episode of Professional Punters, we sit down with Brian Huang, co-founder of Glider. His career began on Morgan Stanley’s rates desk, trading Treasuries, swaps, and swaptions. He then worked at XTX Markets, which he described as a machine-learning firm trading equities, futures, and other linear products around the world.
His route into crypto started at MIT in 2014, when the MIT Bitcoin Project offered undergraduates $100 in Bitcoin and he began trading on GDAX, the predecessor to Coinbase Pro. He now builds Glider, backed by a16z CSX, Coinbase Ventures, Uniswap Labs, and GSR. Brian describes it as a platform for building and running investment strategies with tokenized, onchain assets.
This conversation is about why pension funds trade swaps decades out, how a ML-Driven market maker thinks, how alternative data can become signals, what GameStop taught Brian about regime breaks, and how he thinks about LLM research and execution.
Liability-Matching
Brian’s explanation of rates starts with a pension fund that knows it must pay money out twenty or twenty-five years from now. A Treasury, an interest-rate swap, or a swaption (the right to enter a swap) lets the fund match an asset to that future liability.
The other side rarely arrives on the same schedule. A corporation issues debt when it needs capital, not when a pension plan’s beneficiaries are expected to retire. That mismatch produces flow across the yield curve.
Dealers price this intermediation risk; hedge funds trade whether the curve will steepen, flatten, or invert, and whether fixed rates will move against floating ones.
How XTX Trades
Morgan Stanley and XTX sat at opposite ends of the trade-size spectrum. At the bank, Brian traded with hedge funds and asset managers at institutional size. A million-dollar ticket could be too small to matter; trades ran into tens or hundreds of millions, and sometimes into billions.
XTX lived at the other extreme. During Brian’s tenure, he said, the firm traded essentially every liquid linear market it could reach: equities and futures across North America, Europe, Asia, South America, and South Africa, as both maker and taker. Instead of a few giant tickets, it turned lots of tiny trades into enormous daily volume.
At the shortest horizon, the mechanism can be very simple: S&P futures move in Chicago while a related NYSE stock is still stale. However, XTX’s neural networks takes time to process their inputs, and so the firm tend to focus more ML resources on higher latency signals, while using simpler ones for the lower latency ones.
For example, microwave links are typically involved in cross-exchange arbitrage for speed (the signal does not have to move through a medium). They move data quickly but can lose packets in bad weather, which at times create trading opportunities as the algorithms of certain firms can change in predictable ways.
An ultra-high-frequency portfolio might turn over 10, 20, or 30 times its size in a day, sometimes more. At a trading firm, medium frequency might mean two or three turns a day; at a hedge fund, the same label can mean holding for weeks. More repetitions create more chances to express an edge, which helps explain why the shortest-horizon strategies can post such high Sharpe ratios.
Alternative Data
Prices, volumes, and order books are obvious inputs, so are cross-asset relationships. As such, edge often lies in harder-to-process alternative data.
Brian’s sneaker-reselling business supplied the cleanest example. He could query sales on StockX and eBay. That helped him trade sneakers, but it also gave him an early read on Nike demand. When resale activity fell sharply ahead of a weak Nike earnings report, he put on a large short. He had seen part of the sales picture before it reached the company’s reported numbers.
Lihong recalled a Jane Street mentor arguing that ugly data could contain more alpha because nobody wanted to clean it. Brian immediately qualified the idea: it only works if nobody else already has a clean version. Otherwise the mess is not an edge.
Brian described U.S. equity feeds as relatively centralized and well timestamped; while crypto is scattered across thousands of decentralized exchanges whose timestamps and conventions do not line up cleanly.
Feed a model badly synchronized data and it may learn a relationship that could never have been traded in real time. Brian thinks fragmentation could benefit firms such as Wintermute, while also creating demand for vendors that can deliver clean data.
Neural Net Trading
A neural network does not explain its price the way a simple regression does. Brian compared the problem with asking why an LLM produced a particular answer: billions of parameters can absorb prices, order books, cross-asset moves, and external data without yielding an interpretable story.
So the controls sit around the network. Brian described a three-month rolling training window, with a new model released each week using the previous model’s weights. A security could have separate models for NYSE, Nasdaq, and IEX. The neural-network output then joined simpler components in an aggregate model.
Some components overfit or stop helping in live trading. Regularization techniques that shrinks weak inputs toward zero and real-time monitoring are ways to push those weak pieces out of the aggregate view.
There is also no single t-statistic that certifies a neural network. XTX would run the newly trained model over the previous week and earlier months, then look beyond the headline P&L. Was the distribution unexpectedly skewed? Bimodal? Was one instrument drawing far more volume than it should? Deviation from regularity and prior intuition can expose a problem that the total return hides.
Agentic trading platforms face the same issues that neural nets have. Exposed reasoning can sometimes help with interpretability but the actual decision-making process is still very-much a black box as it is not clear how agents decide to come up with specific lines of reasoning.
Periods of Volatility
A market maker needs flow. When volatility rises, spreads often widen and more people trade. For a hedged firm operating inside a familiar regime, the direction of the move matters less than the volume.
Brian pointed to social-media claims that Wintermute had blown up during a violent crypto move. His guess was the opposite: it was probably one of the firm’s largest P&L days. Selloffs and rallies can both be good trading environments when customers keep crossing the spread.
GameStop supplied the limit. During the 2021 episode, XTX sometimes turned the stock off. Brian wrote a script to scrape WallStreetBets for other tickers that might also need to be disabled. Backtest data could not say much about an unprecedented meme-stock surge. Drawdown limits could stop the system; a human still had to recognize that the regime itself had changed.
Agentic Investing
Brian’s retail advice is blunt: “You want to be buying from a market maker who’s trading every millisecond, and you’re holding for years.” Fighting that market maker in milliseconds means facing better infrastructure and far more repetitions. A retail investor can ignore the micro-movements that the maker is paid to price. That difference in horizon is what makes the trade workable. The market maker may hold for milliseconds while the buyer holds for years.
Different trading firms use frontier models differently. But the suggestion for consumer AI in investing was simpler: let it help with ideas, but keep it away from execution. An ambiguous request for Tesla exposure should not end in an unrelated Tesla meme coin. Asking an LLM whether Bitcoin will rise today, or in the next fifteen minutes, is probably not a productive use of tokens.
Research automation is more credible. Consider workflows that classify job listings to see which firms are hiring, inspect airline seat maps to estimate ticket sales, or use shadows in satellite images to infer oil-storage levels. The data can however become adversarial once the trade is obvious. Companies may post fake openings to frustrate funds reading hiring as a growth signal.
The idea is to start with a hypothesis that requires messy unstructured data, and let AI turn it into cleaner structure data.
Tokenized Stocks
Glider is a way to bundle tokenized stocks and other onchain assets into strategies that run around the clock while users retain custody. The direct-index structure was intended to give users positions in the constituents rather than a pooled fund share.
A semiconductor basket was one example. It was also noted that Glider was working with outside strategy providers (similar to vault curators) and considering factor portfolios, basis trades, and funding-rate arbitrage. Lending and staking may eventually sit inside the same experience.
The ambition is an onchain Fidelity (and less Robinhood): Glider was for people who wanted to “grow their onchain wealth, not gamble it away.”
Professional Punters is presented by Freeport Markets: Freeport lets users trade 24/7, weekends included, with no KYC and up to 200x leverage across stocks, indices, commodities, crypto, rates, and more, with access to pre-IPO exposure in companies like OpenAI & Anthropic. For traders who already follow sharp sources across Twitter, Substack, hedge fund filings, corporate insider activity, and political trading disclosures, Freeport uses AI to read the sources they trust, connect the dots across markets, and surface trade ideas as narratives develop.
About the Host: Lihong Wang is the founder and CEO of Freeport Markets. Before starting Freeport, he traded discretionary semiconductor names at quantitative market-making firms including Jane Street and IMC Trading, building both systematic and discretionary strategies. Lihong graduated from Duke with a degree in mathematics and statistics, and has raised over $2.5M to build Freeport from investors including Y Combinator, Alliance DAO, and Informed Ventures. For more of his long-form research, visit freeportlogbook.substack.com.
Disclaimer: The information provided on TheLogbook (the “Substack”) is strictly for informational and educational purposes only and should not be considered as investment or financial advice. The author is not a licensed financial advisor or tax professional and is not offering any professional services through this Substack. Investing in financial markets involves substantial risk, including possible loss of principal. Past performance is not indicative of future results. The author makes no representations or warranties about the completeness, accuracy, reliability, suitability, or availability of the information provided._
This Substack may contain links to external websites not affiliated with the author, and the accuracy of information on these sites is not guaranteed. Nothing contained in this Substack constitutes a solicitation, recommendation, endorsement, or offer to buy or sell any securities or other financial instruments. Always seek the advice of a qualified financial advisor before making any investment decisions.