When Data Stops Being the Edge: Feature Engineering, LLMs, and What Ordinary Investors Can Learn from Two Sigma

For most of the last two decades, the quant edge was simple: get data nobody else has. Ben Wellington, Head of Complex Feature Engines at Two Sigma, recalls joining the firm in 2007, when data vendors were surprised anyone wanted data on all companies — not just a handful. Today, that same data is table stakes.

The question isn't whether you can buy the data. It's what you do with it once everyone has it.

Wellington frames the modern quant process in three layers: the data you get, the features you build from it, and the forecasts you produce. A "feature" is just a meaningful fact about something you trade — the stock's return over the past week, whether an analyst upgraded it, how much news coverage it's getting. The middle layer — turning raw data into useful features — is where edge increasingly lives. It's also the layer AI is reshaping fastest.

This piece unpacks Wellington's framework from his conversation on Flirting with Models and asks: what from this institutional playbook actually matters for ordinary investors?

What is a feature, really?

A feature is a meaningful fact about something you trade. The stock's return over the past week is a feature. Whether an analyst upgraded it yesterday is a feature. How much news coverage a company gets compared to its own history is a feature.

Wellington's example: you could ask which analyst upgraded the stock, or you could ask does the analyst live in the same city as the company they cover, did they go to the same school as the CEO, how long have they been at this firm. Each question — if you have a reason to believe it matters — becomes a feature. The individual signal from any one is tiny. But two hundred such features, each capturing a slightly different slice of reality, add up to something predictive.

The creativity is in asking the question. The engineering is in answering it cleanly.

How do LLMs change feature engineering?

Wellington's one-liner captures the shift: "Anything can be language now."

Wellington's background is in natural language processing — specifically machine translation. He spent over a decade at Two Sigma studying how text predicts markets. His framing of the shift is precise:

"NLP was about turning language into numbers. With LLMs, I can turn numbers into language. Anything can be language now."

The practical meaning is twofold. First, you can hand an LLM a spreadsheet, a time series of anomalies, or a patent filing, and ask it to describe what it sees — in text. That text output is itself new data. For a firm that has been mining textual signals for fifteen years, an LLM that generates novel text on demand is like discovering a new oil field.

Second, the cost of testing an idea collapsed. Wellington gives the example of studying CEO micro-expressions during earnings calls — blink rates, gaze direction. Five years ago, testing whether CEO blink frequency predicts anything would have meant hiring a dedicated computer vision team and running pipelines for months. With modern LLM tools, you can prototype the idea in an afternoon. The payoff calculation fundamentally changed.

When the effort to test an idea drops from months to hours, your idea queue opens up. Features you'd shelved because they were too expensive become viable overnight.

The democratization trap: when cheap features mean crowded features

But there's a flip side. If you can prototype a feature in an afternoon, so can everyone else.

Wellington describes the risk as AI lowering the entropy of the output — in plain terms, everyone converges on the same answer. If every researcher hands the same public model the same dataset and asks the same question, they all get the same result. That's precisely what you don't want. In portfolio construction, diversity is the whole mechanism — uncorrelated signals combined are stronger than any single signal. If everyone's AI converges on identical features, the diversity disappears, and the edge with it.

This is the core tension: AI makes feature creation cheap, but cheap creation erodes the very scarcity that makes a feature valuable. Wellington's solution is to treat AI as an amplifier of originality rather than a replacement for it. The tool should let a physicist and a computer scientist arrive at different answers to the same question — because their backgrounds, their mental models, and their approaches differ. If the tool forces convergence, it's destroying edge, not creating it.

Collinearity is the graveyard of signals

If you're generating hundreds of features, you'll find many that look new but aren't. Wellington is blunt: "collinearity is the death of a signal."

His example: you build a news sentiment model and think you've found a new signal. But when you check, it turns out the sentiment is just picking up post-earnings stories — it's a roundabout way of measuring earnings numbers you already had. You haven't found anything new; you've found another way to measure something already in your playbook.

The discipline at Two Sigma is that every new feature must be tested against existing factors — market style, industry, and signals already in use — to strip out the overlap. What's left after that, the part that can't be explained by anything you already had, is the only part that matters. If there's nothing left, the feature is dead on arrival, no matter how good its standalone backtest looks.

What is look-ahead bias, and why is it the silent killer?

Wellington raises a concern about LLMs that every backtester should understand: look-ahead bias. Simply put, it's when your model "knows" something it shouldn't — information from the future that leaks into what looks like a prediction.

His example: ask an LLM about Enron, and it immediately associates the company with fraud and collapse. But if you're asking the model a point-in-time question — "what's your outlook on Enron in early 2001?" — the model's knowledge of how the story ends contaminates its answer. It "knows" the future, and that knowledge leaks into what looks like a prediction.

This is not a hypothetical concern. A 2026 paper by Mostapha Benhenda, Look-Ahead-Bench: A Standardized Benchmark of Look-ahead Bias in Point-in-Time LLMs for Finance (arXiv:2601.13770), formalizes exactly this problem. The benchmark measures how standard LLMs — pretrained on web-scale text that includes retrospective analyses — produce artificially inflated backtest results that "evaporate in real-world deployment" once the model's knowledge window ends. The paper evaluates Llama 3.1, DeepSeek 3.2, and specialized point-in-time models, and finds significant look-ahead bias in standard models compared to properly temporal ones.

The lesson is broader than LLMs. Any time your backtest uses information that wasn't available at the time of the decision — whether it's revised economic data, updated corporate filings, or a model that's seen the future — you're not testing a strategy. You're retelling a story you already know the ending to.

Hypothesis-first: the only defense against p-hacking

When features are cheap and experiments are fast, p-hacking is the default outcome. Run enough tests and you'll find "significant" results by pure chance.

This isn't just Wellington's opinion. Campbell Harvey, Yan Liu, and Heqing Zhu documented the scale of the problem in their widely cited 2016 paper "...and the Cross-Section of Expected Returns" (Review of Financial Studies). They catalogued over 300 published factors in top finance journals and argued that a significant portion were likely false discoveries — products of data mining, not genuine economic signals. Their prescription: a new factor should clear a t-statistic above 3.0 (not the conventional 2.0), and researchers must adjust for the number of tests they've run. The more you test, the higher the bar.

Wellington's process at Two Sigma mirrors this. The rule is hypothesis first, test second. Before running any backtest, you state your prior — the economic or financial logic for why this feature should predict returns. Then you test. If the results don't match your prior, you don't flip the sign and call it a discovery. You kill it.

The scientific method is the moat. If you can always invent a story to explain whatever the data shows, you're not doing research — you're telling yourself stories. And in markets, stories are expensive.

The platform vs. the pod

Most quant hedge funds operate as "pod shops" — independent portfolio manager teams competing for capital, each running their own strategies, often duplicating each other's work. Two Sigma runs a shared platform instead.

Wellington's description: it's not one giant portfolio. It's that modelers produce predictions, those predictions feed into larger centralized portfolios, and the aggregation is the firm's edge. Because no modeler controls a specific book, there's no incentive to hoard tools. When someone develops a clever feature or a useful algorithm, they push it back into the shared platform for others to use.

This means every researcher stands on every other researcher's shoulders. A genuinely new feature built by one team becomes a building block for portfolio construction across the firm. The cycle is bidirectional: the platform isn't a static library researchers pull from — it's an evolving ecosystem everyone contributes to.

The contrast matters because it reflects a bet about where edge comes from. A pod shop optimizes locally — each PM maximizes their own book. A platform optimizes globally — the firm maximizes the aggregate. When feature engineering is the edge, global wins, because the value is in the combinations, not the individual pieces.

Knowing when to walk away

Wellington identifies two skills that separate great quants from good ones — and they're both about restraint.

The first: knowing when to stop improving. A model that's 90% done is usually good enough. The last 10% of gains take 90% of the effort, and that effort could have gone into ten new features at 90% capacity instead. For a diversified portfolio, ten different angles beat one perfect one.

The second: knowing when to kill a project. When a backtest signal has no economic logic behind it, walk away — even if the Sharpe ratio looks great. The ability to abandon a losing line of inquiry and redirect energy is more valuable than squeezing marginal alpha from a fading signal.

Why RiskHarvest cares about hypotheses

Two Sigma has hundreds of researchers, proprietary data, and a shared platform built over two decades. Most investors have none of that. But that's exactly why the principles matter — they're the part that translates.

At RiskHarvest, we start from the same place Wellington describes: the hypothesis comes first. Before we backtest a strategy, we ask why it should work. What risk premium does it capture? What economic mechanism drives the return? If we can't answer that in plain language, the backtest doesn't matter — a good Sharpe ratio without a reason is just noise.

Then we check: is this signal actually different from what we already have, or is it just another way of measuring the same thing? Is the data clean — or has the future leaked in? And if the answer is "this doesn't hold up," we walk away. No amount of optimization fixes a signal that was never real.

This is why we focus on premia that are well-documented, replicable, and accessible to ordinary investors through low-cost instruments. Not because feature engineering is beyond retail — but because the foundation has to come first. Harvest the premia that are already there, understand why they exist, and build confidence in the process. Once that foundation is solid, the more creative work of feature engineering has a base to stand on.

The institutions that are good at this aren't good because they have data you can't get. They're good because they have discipline — sharper hypotheses, stricter checks, and the willingness to abandon ideas that don't survive scrutiny. That discipline isn't proprietary. It's the one thing that scales down to any account size.

Educational only. Not financial advice.