Sparse Signals: When the Lasso Found Real Alpha in One-Minute Returns

Here is a question almost nobody had asked properly.

Forget factors. Forget economic theory. Forget everything you think you know about what predicts a stock's return. Just take the recent returns of every single stock in the market, all of them, thousands of candidate predictors, and ask: does any subset of those predict this stock's return over the next minute?

It is an absurd question in the traditional framework. You have thousands of candidate predictors and a short estimation window. Ordinary regression cannot handle it. And there is no economic story telling you why the return of a random mid-cap industrial should predict the return of a bank one minute from now.

Alexander Chinco, Adam Clark-Joseph and Mao Ye asked it anyway, using the lasso, and the answer they got is one of the more interesting results in financial machine learning, because of what they found the machine had found.

The problem: theory tells you where to look, and it may be looking in the wrong place

The traditional approach to return prediction starts with an economic idea. You believe that value predicts returns, so you build a value signal and test it. The economics comes first, the statistics second.

This is disciplined, and it protects you from data mining. It also has a blind spot: it can only find predictors that a human thought of in advance.

Suppose the real predictors of a stock's next-minute return are strange, specific, and fleeting. Suppose that right now, for the next hour, this particular bank's returns are being predicted by that particular industrial's returns, because of a supply-chain link, or a shared exposure, or a piece of news that is diffusing slowly through the market. Suppose that relationship exists for an hour and then disappears forever, replaced by a completely different one.

No economic theory would ever tell you to look there. And no fixed factor model could capture it, because the relationship is not stable enough to be a factor. It is here, and then it is gone.

The key idea via analogy: a searchlight, not a map

The authors' method is direct. At each point in time, they run a lasso regression to forecast a stock's return over the next minute, using as candidate predictors the lagged returns of the entire cross-section of stocks. Thousands of candidates. Then they roll forward, and do it again, and again.

The lasso is exactly the right tool for this, for one reason: it produces sparsity. Out of the thousands of candidate predictors, it selects a small handful and sets everything else to exactly zero. It does not need you to tell it where to look. It searches everywhere and reports back a short list.

The analogy: a traditional factor model is a map, drawn in advance by economists, showing where the treasure is supposed to be. The lasso is a searchlight, sweeping across the entire market, stopping wherever it catches a glint. It has no theory. It has no idea what a bank is or what a supply chain is. It just looks.

And when the authors rolled this forward through the data, the searchlight kept finding things. Out-of-sample predictive accuracy improved, and the forecast-implied Sharpe ratios were meaningfully better than the benchmarks. The machine was finding real, exploitable structure.

The finding that makes the paper important

Anyone can build a model that appears to predict. The reason this paper landed in the Journal of Finance is what the authors did next: they went and looked at what the lasso had selected, and asked whether it made any sense.

They report three properties of the predictors it found.

They are sparse. At any moment, only a handful of the thousands of candidates matter. The signal is concentrated, not spread thinly across everything.

They are short-lived. The predictors that matter this hour are not the predictors that matter next week. The relationships appear and vanish. This is why no fixed factor model could ever have captured them: by the time you have identified one and written a paper about it, it is gone.

They are unexpected. The stocks the lasso picks as predictors are not the ones an economist would have nominated. They are not obviously in the same industry, not obviously linked.

And then the payoff. The authors check what is happening to the stocks that the lasso selects as predictors, and find that they tend to be stocks with news about fundamentals.

That is a genuinely satisfying result, and it changes the interpretation of the whole exercise. The lasso, working with no economic knowledge whatsoever, using nothing but a statistical criterion, was detecting the arrival of news and tracking how information diffuses across the market. When something real happens to one company, it takes time for the implications to be priced into related companies, and the lasso was catching that lag, live, without being told what news is or which companies are related.

The statistical rule found the economics on its own. That is the paper's central claim, and it is a rebuttal to the standard objection that machine learning in finance is just data mining with better marketing.

Why it mattered

• It is a rare case of machine learning finding alpha and then explaining itself. Most financial machine learning papers show improved prediction and stop. This one goes further, showing that the mechanically selected predictors correspond to a comprehensible economic event, which is what turns a data-mining result into a finding.
• It demonstrated that useful signals can be sparse and unstable. The traditional research programme hunts for stable predictors, because a predictor that vanishes cannot be published as a factor. This paper suggests a large amount of genuine predictability lives in relationships that are fleeting by nature, and that the entire methodology of looking for stable factors is structurally blind to it.
• It made a serious case for the lasso as a discovery tool. Not as a way of shrinking a model you already believe in, but as a way of searching an enormous space that no human could search, and coming back with a short list worth examining.
• It is a template for honest machine learning research. Predict, then interrogate what the model learned, then check whether it corresponds to something real. That sequence is what separates this from the vast literature of unfalsifiable backtests.

The honest limitations

• This is high-frequency prediction, and profit is a separate question. The horizon is one minute. At that horizon you are fighting the bid-ask spread, exchange fees, and the fact that you must actually get filled. Forecast-implied Sharpe ratios are a statistical construct. Whether a real trading operation could capture these signals, net of costs and market impact, is not something the paper settles.
• Fleeting signals decay, and this one is now published. A relationship that lives for an hour is exactly the kind of thing that gets competed away once a large number of well-resourced participants are hunting it with the same tool. The value of this paper to a practitioner in 2026 is probably more methodological than directly tradeable.
• The lasso's selections are unstable, and that is uncomfortable here. When candidate predictors are correlated, the lasso picks one somewhat arbitrarily. The paper's story depends on the selected predictors being meaningful, and the fact that the lasso would have picked a different one with slightly different data complicates that interpretation.
• The news finding is an association, not a mechanism. Showing that selected predictor stocks tend to have news is a strong and welcome piece of evidence, but it does not establish exactly how the information diffuses or why it takes the specific paths it takes.
• You cannot easily rerun this at home. The data requirements, minute-level returns across the entire cross-section over a long period, are substantial, and the rolling lasso is computationally heavy.

The one-line takeaway

Chinco, Clark-Joseph and Ye pointed a lasso at the entire cross-section of lagged stock returns with no economic theory to guide it, found that it produced genuinely better one-minute forecasts, and then showed that the strange, fleeting predictors it selected were stocks with fresh news, meaning the algorithm had discovered, on its own, the slow diffusion of information across the market.