Brett Harrison

Brett Harrison@BrettHarrison · 14 天前观点输出

Hedge funds will eventually provide inference to model routers for free, or even pay to do so, in order to get a first look at prompt contents. Payment-for-order-flow in the age of compute.

901051957205.9K

AI：延续个人公开立场，强调长期主义、执行效率与行业方向判断。

Brett Harrison@BrettHarrison · 15 天前政策影响

The first reaction to finding a high-Sharpe strategy in simulation should be to assume there’s a mistake. Markets are adversarial, data is messy, and small modeling errors compound into misleading results. Common mistakes in training systematic trading models and some solutions: Omitting fees and slippage: Nearly all equities, options, futures, and digital asset exchanges charge higher fee amounts for removing liquidity than for adding liquidity. HFT strategies that hold positions for only seconds or minutes frequently see their expected edge disappear once realistic fees are applied. Another common mistake is assuming a liquidity-taking order is fully filled at the best bid/ask instead of walking the book, consuming successive price levels up to available size. This both systematically underestimates cost and overestimates the capacity of the strategy. Accurately modeling fees and slippage requires building automated reconciliation between predicted costs and exchange-reported fees and fill prices. In-sample contamination: Statistical arbitrage strategies that fit model parameters to historical data need to avoid evaluating performance on the period used for training. The naive method of splitting data is to hold back the most recent days from the training set, which may result in training the strategy on a different market volatility and momentum regime than the recent past. Simple K-fold cross validation, a common fix from machine learning, is inappropriate for time series as this method leaks future information into the past. The most successful data splitting techniques involve purged K-fold, walk-forward optimization, or combinatorial purged cross-validation. Incompatible clocks: Backtesters that reference timestamps based on the market data capture machine’s clock overestimate achievable fill ratios. A partial solution is to refer to exchange timestamps for order book events to determine whether liquidity still exists when orders are sent. However that solution is also unreliable as exchanges have wide distributions of latencies between matching engine actions and book update dissemination. A robust heuristic is to re-run the backtest under a range of artificial order-submission delays and examine the sensitivity of both fill ratio and P&L. Survivorship bias: The choice of instrument universe itself introduces survivorship bias. For example, choosing the constituents from today’s S&P 500 index systematically selects historical winners only, as those names that underperformed have already been excluded from the basket. One solution is to record point-in-time universes to use for future historical studies, and another is to fix today’s universe and implement a walk-forward simulation until enough data has been collected.

7111095.9K

AI：偏政策推进，强调合规落地与规则明晰，对监管主线更敏感。

Brett Harrison@BrettHarrison · 16 天前政策影响

Excellent and comprehensive discussion of the critical market opportunities in compute derivatives. Appreciate Architect’s inclusion as US exchanges approach regulatory approval and begin building robust futures markets in compute. https://t.co/VToq7ZKvMf

588612.4K

AI：偏政策推进，强调合规落地与规则明晰，对监管主线更敏感。

Brett Harrison@BrettHarrison · 16 天前政策影响

Can inference costs be commoditized? On OpenRouter there’s an 11x spread between the cheapest and most expensive provider of a single model, DeepSeek V4 Flash. Other relevant data points on open-weight inference prices, speeds, availability: • Baidu serves DeepSeek V4 Flash at $0.049 per million tokens and at 124 tokens per second. 26 of the 30 other providers are both more expensive and slower, so it’s not a tradeoff of speed vs cost. • Across the 117 models on OpenRouter with three or more providers, the median price spread between the cheapest and priciest host is 2.2x, the most extreme is 14.4x (DeepSeek v3.2). • Among all endpoints serving the above, 37% of endpoints are strictly dominated by cost, speed, and uptime. The number goes up to 52% if considering only cost and speed.

988211.4K

AI：偏政策推进，强调合规落地与规则明晰，对监管主线更敏感。

Brett Harrison@BrettHarrison · 19 天前政策影响

Accurately pricing derivatives when underlyings are closed is a universal source of edge in market making, from tier-1 firm graybox ETF trading to newer operations in 24-hour prediction markets and perpetual futures. Some widely applicable pricing strategies: Related futures and currency moves. When a derivative’s underlyings are closed, market makers use other instruments as proxies that have sufficient beta to the derivative. Examples for a US ETF on Japanese stocks: the USD/JPY pair trades 24/7 through a variety of FX ECNs, interdealer platforms, and futures exchanges. Nikkei 225 index futures also trade on multiple global derivatives exchanges during extended hours. Moves in these assets during the Japanese night will on average predict the opening prints of individual stocks listed on JPX. Home market index moves. An ETF moves with a non-trivial correlation factor to other US names simply because it’s a US-listed security. The effect is easily observable during heightened volatility. It’s a common industry saying that in a market-wide selloff, all correlations go to 1. Both narrow- and broad-based indexes of US names explain some of the signals in mid-frequency alphas. US index returns comprise a small but meaningful component of derivatives’ multifactor beta models. News. A key requirement for pricing derivatives on foreign stocks is processing local news and earnings releases that relate either to the particular stocks or relevant stocks in the same sector. Quantitative trading firms use automated translation tools to process local foreign-language news and suggest idiosyncratic adjustments to traditional factor models. Recent advances in LLM-based NLP have made sentiment analysis viable for blackbox trading systems to react instantaneously to news-based signals. Microstructure. High-frequency trading firms successfully and counterintuitively price derivatives by ignoring the underlyings’ characteristics. Firms extract short-term alphas from order book characteristics, recent returns, microprices, and other microstructure features. From the perspective of a reinforcement-learning non-linear model trainer, unlabeled feature sets and time series data result in ETFs, ADRs, and common stocks being treated as mathematically equivalent. Based on the above, here is a rough model of how large market-making firms predict mid-frequency returns when traditional underlying markets are closed: ΔP = β_FUT · ΔFUT + β_FX · ΔFX + β_IND · ΔIND + [News Term] These firms also heavily prepare for circumstances under which this model fails. FX and futures have idiosyncratic moves, home-market betas can break during regime shifts, news sentiments still have positives, and microstructure signals decay fast. The market shift underway in 24/7 traditional derivative and prediction contract trading will test the model’s longevity for market making and statistical arbitrage going forward.

71025018.8K

AI：偏政策推进，强调合规落地与规则明晰，对监管主线更敏感。

Brett Harrison@BrettHarrison · 20 天前政策影响

AI has broadly become synonymous with LLMs, but most AI/ML models used in quantitative trading are decades old. Boosted trees and PCA still account for more alpha than anything built on transformers. More detail on several ML techniques in quant trading: Gradient boosted trees and random forests: A standard strategy in short-horizon trading is to start with an intuitive set of microstructure features and train a non-linear model on them. Random forests provide a baseline ensemble that’s hard to overfit, while gradient boosted models fit fainter structures at the risk of overfitting. Features can include microprices, characteristics of the present order book, recent returns of related instruments, momentum signals, and orderflow directionality. Pre-transformer natural language processing: Throughout the 2010s, NLP PhDs were a sought-after resource for quantitative hedge funds. A variety of techniques such as Loughran-McDonald finance-specific sentiment dictionaries, TF-IDF and bag-of-words classifiers, and LDA topic modeling were used to extract sentiment from news, earnings reports, SEC filings, and FOMC minutes. Transformers are more generalized tools for language processing but some of the older methods in the toolkit are still used for their simplicity, predictability, and cost-effectiveness. Dimensionality reduction: A statistical arbitrage researcher may have thousands of correlated instruments and only a few years of clean historical data, so the covariance matrix has more parameters than observations to estimate them from. Principal component analysis is the standard remedy. A handful of leading components explain most of the variation in a universe of equities, and hedge funds will frequently neutralize the leading factors and trade the residuals. Autoencoders generalize the same idea to nonlinear structure, but PCA persists because its output is stable, interpretable, and cheap to recompute. Computer vision: From around 2010 to 2017, hedge funds gained a large informational edge by building automated pipelines to extract alternative data from satellite imagery. The models were unremarkable by research standards, mainly derivatives of convolutional neural networks. Some of the most famous applications include counting parked cars in retail lots, gauging construction progress at industrial sites, estimating crop yields from near-infrared reflectance rather than visible light, and reading crude inventory off the shadow cast by a floating tank roof. The trade worked best during a time when imagery was scarce and the engineering was novel. Today image-based alternative data is sold by third-party vendors for relatively inexpensive consumption. The large memory footprint, slow inference, and black-box nature of LLMs make them a poor fit for alpha generation at trading firms. Hand-tuned models based on intuitional feature selection are still the norm. For niche industries like quantitative finance, there is ample room for competition with generalized frontier models.

015291

AI：偏政策推进，强调合规落地与规则明晰，对监管主线更敏感。

Brett Harrison@BrettHarrison · 21 天前政策影响

The SEC’s Regulation Crypto Assets will create needed pathways for digital securities primary issuance, but regulated secondary markets will remain impossible without the agency rescinding Rule 611, the Order Protection Rule. More on this rule and its relevance to crypto: Rule 611 under SEC Regulation NMS prevents any venue for US equities or equity options from matching trades at a worse price than a protected quote on another venue. A protected quote is a round-lot firm order at top-of-book on a lit national securities exchange. Before 611 was established, floor brokers with access to quotations on multiple exchanges could provide inferior pricing to clients. The rule accelerated the migration to fully automated execution and led to today’s competitive ecosystem of 17 lit equities exchanges, 18 lit options exchanges, and over 80 darkpool ATSs and single dealer platforms. Major market participants have grown more critical of 611 in recent years. To comply with order routing rules, exchanges and broker-dealers pay market data and connectivity fees to the growing list of NMS exchanges. Large orders that can’t be filled on one exchange get broken up slowly and suffer price impact before routing is complete. The protections are less needed in an environment where HFTs already maintain price equilibria between exchanges through arbitrage. The new proposal for SEC Regulation Crypto Assets would create a safe harbor for primary issuance of digital securities, but doesn’t address how secondary markets can develop in light of the Order Protection Rule. An ATS or on-chain venue listing a token could not comply with the rule without NMS venues listing the asset, establishing a canonical NBBO, and facilitating on/off-chain hybrid routing. The SEC proposed rescinding the Order Protection Rule in June. Doing so would provide a clearer path for digital securities trading, but introduce significant disruption and unknown impact on equity and equity options markets.

812714.9K

AI：偏政策推进，强调合规落地与规则明晰，对监管主线更敏感。

Brett Harrison@BrettHarrison · 22 天前观点输出

@AlpacaHQ Congrats @iyoshyoshi @umitanuki and team! おめでとう

100259

AI：延续个人公开立场，强调长期主义、执行效率与行业方向判断。

Brett Harrison@BrettHarrison · 22 天前赛道影响

Another consequence of inference unit economics improving vs training: it’s profitable to run inference clouds on scrappy datacenters like stranded gas pads, old crypto mines, or telco closets. Any powered shell with decent chips can be turned into a token-refining revenue gen. https://t.co/3M9AAdfyWK

102517814.9K

AI：偏平台经营与生态扩张，强调交易平台和应用入口的长期位置。

Brett Harrison@BrettHarrison · 23 天前政策影响

I’ve built HFT/exchange systems in three languages that sit at very different points on the safety/performance/ergonomics spectrum: OCaml, C++, and Rust. Observations and a practical comparison from production use: OCaml • Strong static typing and algebraic data types make illegal states almost impossible to represent. Protocol parsers, order book models, and risk checks are self-documented by types and function signatures. • The default garbage collector is a non-starter for predictable tail latencies. There are workarounds using disciplined zero-allocation techniques: custom core libraries, careful use of the [@ inline] and [@ unboxed] attributes, and a culture that treats any heap allocation in the hot loop as a bug. • Functional programming forces strategy writers to think in pure transformations of market data to fair values to orders. This style enables multiple forms of unit testing and makes offline model training pipelines easier to implement without major code changes. • The ecosystem for kernel-bypass networking and FPGA offload is thin; performant OCaml requires writing a fair amount of C interop or custom FFI. Compile times and binary size can also become painful at scale. • Engineering effort on type safety is a replacement for future engineering effort on debugging. We used to half-joke, “if it compiles, it works.” C++ • The most commonly used language in HFT because it offers maximum control. The writer can pinpoint every cache line, allocation, and branch. When the last microseconds matter, few languages can compete. • Industry standard for exchange connectivity, FPGA drivers, and shared-memory IPC. Virtually every low-latency NIC vendor’s API is written for it. • The cost is constant vigilance. Use-after-free, data races, and subtle ABI mismatches are permanent time sinks. Even with modern C++ features such as concepts, spans, and std::atomic memory orders, the cognitive load stays high. • Template metaprogramming offers some of the type-level guarantees of OCaml or Rust, but the error messages and compile times can be inscrutable. • Most of the “C++ is faster” advantage narrows when factoring in the engineering time spent hunting undefined behavior. Rust • Ownership and borrowing give the memory-safety guarantees of a garbage-collected language with the performance profile of C++. Zero-cost abstractions allow HFT system authors to write zero-allocation, lock-free, cache-aware code and the compiler will refuse to allow introduction of data races. • Excellent for the modern low-latency stack: DPDK/AF_XDP bindings, io_uring, shared-memory rings, and safe wrappers around exchange binary protocols. • The type system is expressive enough to encode many of the same invariants of OCaml, such as phantom types, newtypes, and state machines, without the garbage-collection tax. • The borrow checker still occasionally forces awkward redesigns of hot-path data structures, async/await is not yet as mature for ultra-low-latency work as plain threads/ busy-polling, and the ecosystem for certain HFT-specific libraries lags C++. • It takes time for the ownership model to click, after which mental bandwidth can be dedicated to strategy and latency. It’s much more difficult to build a high-frequency trading system that retains correctness and readability as it scales than it is to build one that’s fast. Good taste and developer culture can coalesce around a particular language, but the best teams possess these qualities independently, and can adapt freely and agilely to new technologies as they become available.

384780750.9K

AI：偏政策推进，强调合规落地与规则明晰，对监管主线更敏感。