Quick answer: Quant funds don't read the news the way you do. They convert it into numbers. Algorithms parse machine-readable feeds (Bloomberg, Dow Jones, Refinitiv) within milliseconds, score each story for sentiment and relevance, and feed those scores into models as predictive features alongside price, volume, and other factors. Some chase speed (latency arbitrage), most chase signal (sentiment factors with multi-day horizons). The edge fades as more capital piles in, which is called alpha decay. The interesting part for 2026: tools like NowNews now give retail investors a watered-down but real version of the same sentiment-and-impact scoring that used to cost six figures a year.
Want to see what news-as-data actually looks like on a chart? You can start a free 7-day NowNews trial and watch news markers, sentiment, and price line up on the same screen. No card required.
There's a myth that quant funds have some magic crystal ball. They don't. What they have is a pipeline. A boring, relentless, very well-engineered pipeline that turns words into features and features into positions. Once you see the pipeline, the mystery mostly evaporates.
I'll walk through how that pipeline actually works, from the moment a headline hits the wire to the moment a model adjusts a position. We'll cover the data feeds, the NLP scoring, the strategies (fast and slow), the academic numbers behind it, and the part nobody likes to talk about: how fast these edges die. And at the end, the honest framing on what a retail investor can and can't replicate in 2026.
What "using the news" really means to a quant
Start with the obvious gap. A discretionary investor reads an article, forms a view, maybe trades. A quant fund never does that. The fund treats news as a structured data stream, the same way it treats price ticks or order-book depth.
So the first job isn't analysis. It's conversion. Take an unstructured blob of text and squeeze it into fields a model can ingest: which company is this about, is it positive or negative, how novel is it, how relevant, what's the event type. RavenPack, one of the longtime vendors on quant desks, classifies news into nearly 6,900 categories and attaches scores like an Event Relevance Score and an Event Similarity (novelty) score to each story. (Source: RavenPack.)
Why bother scoring novelty? Because the tenth article repeating the same earnings miss carries almost no new information. The model wants the first signal, not the echo. This is the part that surprises people the most. A huge amount of quant news work is about ignoring news, specifically the redundant 95% of it.
Once you've got numbers, everything downstream is familiar quant machinery. You backtest. You combine the sentiment feature with other factors. You size positions by conviction and risk. The news is just another column in a very wide spreadsheet.
The data pipeline, step by step
Here's the rough flow most systematic shops use, simplified but accurate:
- Ingest. Machine-readable news feeds arrive from providers like Bloomberg, Dow Jones Newswires, Refinitiv (LSEG), and AlphaFlash. These are built for machines, not humans, with tagged tickers and timestamps down to the millisecond.
- Parse and tag. NLP extracts entities (which companies, which people, which assets) and event types (earnings, M&A, guidance, litigation, downgrades).
- Score. Each story gets a sentiment value, a relevance score, and a novelty score. Some pipelines add a confidence or "honesty" check, comparing narrative tone against the hard numbers in a filing.
- Aggregate. Individual story scores roll up into an asset-level or sector-level sentiment indicator over a chosen window (intraday, daily, two-week).
- Model. The aggregated sentiment feeds a predictive model as a feature, blended with price momentum, volatility, fundamentals, positioning, and so on.
- Execute. The model outputs target positions. Execution algorithms work the orders to minimize market impact.
Notice step 6. The trade is almost an afterthought. By the time a position changes, the interesting decisions were all made upstream, in how the text got scored and combined.
News providers have even started doing step 2 and 3 for the funds, selling pre-extracted signals so a trading firm saves the milliseconds it would take to parse the article itself. (Source: Federal Reserve IFDP working paper.) That's how compressed the timeline has become at the fast end.
The two families of news strategy: speed vs signal
People lump "news trading" into one bucket. It's really two very different games.
Family 1: latency arbitrage (the speed game)
This is the milliseconds world. The idea is brutally simple. Information moves prices. If you can read and act on a release a few milliseconds before everyone else, you capture the move.
The academic evidence on just how fast this is got is wild. Chordia et al. (2018) and Hu et al. (2017) found that security prices begin reacting to scheduled macroeconomic news within the first five milliseconds of release. Five. (Source: MIT Sloan; Federal Reserve.) For context, a human blink takes roughly 100 to 400 milliseconds. The fast end of this market is over before you've physically registered the headline exists.
This game is capital- and infrastructure-intensive. Co-located servers next to exchanges, dedicated fiber, the works. It's effectively closed to anyone without a serious tech budget. Worth being honest about that up front: no retail tool, NowNews included, plays in the millisecond arms race. Nobody should pretend otherwise.
Family 2: sentiment factors (the signal game)
This is where most systematic news alpha actually lives, and it's far more accessible conceptually. Instead of racing the clock, you build a factor: a measurable, persistent relationship between news sentiment and future returns over hours or days.
The numbers here are concrete. In RavenPack's APAC case study, news-sentiment strategies on short-to-medium horizons produced Information Ratios close to 4.0 and annualized returns above 20% (before the usual caveats about costs and capacity). (Source: RavenPack research.) In FX, a contrarian one-day strategy on EURUSD generated about 7.2% per year with an Information Ratio of 0.75 after transaction costs. (Source: RavenPack.) And aggregated news sentiment showed 20-27% correlation with two-week forward returns on the DJIA in their work. (Source: RavenPack.)
Those are not get-rich numbers. An Information Ratio near 4 is excellent but applies to a specific period and market. The point isn't the exact figure. It's that the relationship is real, measurable, and slow enough that you don't need a server in the exchange basement to act on it.
This second family is the one retail investors can genuinely learn from. More on that later.
If you'd rather not build a sentiment model from scratch, the NowNews 7-day trial lets you see sentiment scoring and impact classification applied to real assets, so you can judge for yourself whether news-as-data fits how you invest.
How NLP turns a headline into a number
Let's get concrete about the scoring, because this is the heart of it.
Say a wire prints: "MegaCorp cuts full-year guidance, cites soft demand." A human reads dread. An NLP model does several things at once:
- Entity resolution: this is about MegaCorp (ticker), not some subsidiary with a similar name.
- Event classification: this is a guidance cut, a known high-impact event type.
- Polarity: negative sentiment, strong.
- Relevance: is MegaCorp the subject or just mentioned in passing? Subject. High relevance.
- Novelty: has this been reported in the last few hours? If no, high novelty, the signal counts more.
Modern pipelines increasingly use large language models for this, not just keyword dictionaries. A 2025 arXiv paper on combining LLMs with reinforcement learning for sentiment-driven quantitative trading is one of many recent attempts to push beyond crude positive/negative scoring into context-aware reading. (Source: arXiv 2510.10526.) The frontier is models that understand that "beat estimates but guided down" is a mixed signal, not a clean positive.
Here's a nuance retail investors miss. Sentiment polarity alone is a weak signal. The strong signals come from combining polarity with novelty and with how the price already moved. A negative story on a stock that already dropped 8% is often a fade candidate, not a short. Quant models encode that. Your gut usually doesn't.
The honesty problem: narrative vs data
One underappreciated use of news analytics is contradiction detection. Companies craft narratives. The hard numbers in a filing sometimes say something different. Spotting the gap between the story management tells and what the data shows is genuinely valuable, and it's exactly the kind of pattern a model can flag at scale.
This is where the line between "news analysis" and "document analysis" blurs. SEC filings (10-K, 10-Q, 8-K) are news to a quant model, just slow-moving, structured news. RavenPack and others have long mined filings for tone shifts, hedging language, and risk-factor changes year over year.
NowNews builds a version of this into its Deep Analysis feature, with what it calls honesty signals: flagging where the narrative tone in a document diverges from the underlying figures. It's not the institutional-grade contradiction engine a Two Sigma runs internally. But it's the same idea, packaged for someone who doesn't have a data science team. AlphaSense and Hebbia attack the same problem from the enterprise side, with deeper document coverage and a price tag to match.
Alternative data: news is just the gateway drug
News was the first alternative dataset to go mainstream. It opened the door to everything else. Credit-card spend, satellite imagery of parking lots, app-download trends, shipping data, web-scraped pricing. Funds now blend dozens of these.
The pattern across all of them is the same and it's sobering. A genuinely proprietary dataset typically provides something like 12 to 24 months of meaningful alpha before widespread adoption commoditizes it. (Source: industry analysis via search.) After that, the edge erodes because everyone's trading the same signal.
Which brings us to the thing nobody at a marketing booth wants to discuss.
Alpha decay: why these edges die
Here's the honest framing. Every news signal is a depreciating asset.
The mechanism is simple. A signal that generated, say, 2% annual alpha a decade ago might generate 0.5% today, because dozens of funds now trade the identical pattern. The crowd arbitrages away the very inefficiency that made the signal profitable. (Source: industry analysis via search.)
This is why quant funds are never "done." They're on a treadmill: find an edge, exploit it, watch it decay, find the next one. The fancy term is the edge creation-and-degradation cycle. The unfancy version: today's secret is next year's commodity.
It also explains why the retail democratization story has a real basis. The sentiment signals that gave funds an edge years ago have decayed enough that vendors will now sell them, or build consumer products around them, without cannibalizing their crown jewels. When something becomes a $15-a-month app feature, it's usually because the institutional alpha already got squeezed out of it. That's not a knock. It just means retail tools deliver context and speed-of-understanding, not a hidden money machine.
How AI tools change this for non-quants
So where does that leave a regular active investor in 2026? In a genuinely better spot than five years ago, with realistic expectations.
You can't run latency arbitrage. You can't license a full RavenPack feed (institutional pricing runs into five and six figures annually). What you can do is borrow the second family's logic: treat news as scored data, watch sentiment alongside price, and filter for impact instead of drowning in headlines.
A few tools occupy this middle ground:
- NowNews focuses on the retail price point. Its Pulse Signal overlays news markers, 24-hour sentiment, and a pressure indicator directly on an asset chart, so you can see the news-price relationship the way a quant sees it in a backtest. Its Impact Feed filters news by estimated impact and direction, which is the consumer version of relevance scoring. Plans run €14.99, €24.99, and €59.99 a month. One honest limitation: it's desktop-only for now, with a mobile app described as coming soon.
- AlphaSense and Hebbia sit at the enterprise tier. Far deeper document coverage, built for analysts and funds, priced accordingly.
- Bloomberg Terminal remains the institutional default for machine-readable news and far more, at roughly $30k a year per seat.
- Stock Titan and Benzinga Pro target active retail traders with fast news and alerts, lighter on the sentiment-as-data angle.
- ChatGPT and Perplexity can summarize and discuss news well, but they're general assistants, not real-time scored feeds. They don't watch a ticker for you or score every story on impact. Different tool, different job.
The realistic pitch: tools like NowNews won't make you a quant. They give you a quant-flavored lens. You still decide what to do with what you see. That's the part no model hands you.
Curious how the news-as-data lens feels in practice? You can try NowNews free for a week and see your watchlist through sentiment and impact scores instead of an endless headline scroll.
A worked example: one headline, two funds, three reactions
Let's make this tangible. Imagine a single 8-K hits the wire at 9:42 a.m.: a mid-cap software company announces its CFO is resigning "to pursue other opportunities," effective immediately.
A discretionary investor sees the headline, feels a twinge of worry, and maybe googles around for ten minutes before deciding anything. By then it's over.
Now watch the two quant families.
The latency fund doesn't even interpret much. Its model has a prior: abrupt CFO departures, especially "effective immediately" ones, skew negative. The NLP tags the event type, scores it negative with high relevance and high novelty, and the model trims or shorts within milliseconds of the wire timestamp. It's not betting on the story. It's betting on the statistical tendency of stories like this, executed before the crowd finishes reading.
The sentiment-factor fund plays a slower hand. It doesn't trust a single headline. It waits for the sentiment aggregate to build over the next hours: follow-on coverage, analyst notes, the tone of the company's own statement, whether the language hedges ("mutual decision") or alarms ("immediate," "internal review"). If aggregated sentiment stays sharply negative and novel, the position adjusts on a one-to-three-day horizon. This fund is essentially asking: is this a real signal or noise that'll mean-revert by Thursday?
And here's the retail angle. With a tool like NowNews, you wouldn't catch the millisecond move, but you'd see the news marker appear on the chart, the sentiment swing in the 24-hour window, and an Impact Feed entry classifying the event's likely direction. That's the sentiment-factor fund's information set, minus the automated execution. You still pull the trigger yourself. But you're no longer the discretionary investor googling in the dark. You're working from scored data, just slower and cheaper than the fund.
Three reactions to one headline. Speed, signal, and lens. Most retail tools, NowNews included, live firmly in that third category, and that's the honest place to set expectations.
A quick reality check on what quant news trading is not
Three myths worth killing:
- It's not mind-reading. Models react to information faster and more consistently than humans, but they don't predict surprises. An unexpected event is unexpected to the algo too.
- It's not infallible. Fast news analytics can amplify mistakes. Garbled or misclassified headlines have caused flash moves. INSEAD researchers have written about the systemic risks of everyone trading the same fast signal at once. (Source: INSEAD Knowledge.)
- It's not free money. Between alpha decay, transaction costs, and capacity limits, even good signals deliver modest, hard-won returns. The 20%+ figures from case studies are best-case, specific-period, specific-market results. Treat them as illustrations, not promises.
FAQ
How do quant funds use news in their trading models?
They convert news into structured data. Machine-readable feeds are parsed by NLP, each story is scored for sentiment, relevance, and novelty, and those scores become features in predictive models alongside price and volume. The model, not a human, decides how to adjust positions. Tools like NowNews apply a simplified version of this scoring for retail investors.
What is machine-readable news?
It's news formatted for algorithms rather than people, with tagged tickers, event types, and millisecond timestamps. Providers include Bloomberg, Dow Jones Newswires, Refinitiv (LSEG), and AlphaFlash. Funds use it because it can be parsed and scored in milliseconds, far faster than a person could read.
Can news sentiment actually predict stock returns?
To a measurable degree, yes, over short to medium horizons. RavenPack research found aggregated news sentiment correlated 20-27% with two-week forward DJIA returns, and APAC sentiment strategies posted Information Ratios near 4.0 in their studies. The relationship is real but modest, period-specific, and weakens as more capital trades it.
Do quant funds trade in milliseconds on news?
The fastest ones do, in a strategy called latency arbitrage. Prices begin reacting to scheduled macro news within roughly five milliseconds of release, per academic studies. But most systematic news alpha lives in slower sentiment-factor strategies that act over hours or days, which is the part regular investors can actually learn from.
What is alpha decay in news trading?
It's the gradual weakening of a signal as more funds trade it. A pattern worth 2% annual alpha a decade ago might be worth 0.5% today. Proprietary datasets typically deliver meaningful edge for only 12 to 24 months before commoditization. It's why quant funds constantly hunt for new signals.
What is RavenPack and why do quant funds use it?
RavenPack is a longtime news-analytics vendor that scores stories for sentiment, relevance, and novelty across nearly 6,900 event categories. Quant desks license it as a ready-made signal source so they don't have to build NLP pipelines from scratch. Its pricing targets institutions, not retail.
Can retail investors copy what quant funds do with news?
Partly. You can't run latency arbitrage or afford institutional feeds, but you can adopt the core logic: treat news as scored data, watch sentiment next to price, and filter by impact instead of volume. Retail tools like NowNews, Stock Titan, and Benzinga Pro bring pieces of this to a consumer price point.
Does NowNews help analyze news like a quant fund?
It applies a retail-grade version of the same ideas. Pulse Signal overlays news markers and sentiment on a price chart, Impact Feed filters news by estimated impact and direction, and Deep Analysis flags where a document's narrative diverges from its numbers. It's not an institutional quant platform, and it's desktop-only for now, but it gives you the same lens at €14.99 to €59.99 a month.
Is ChatGPT or Perplexity good for news-based trading?
They're useful for summarizing and reasoning about news, but they aren't real-time scored feeds. They won't continuously monitor a ticker or assign impact scores to every story. For systematic news monitoring, a purpose-built tool fits better. For one-off questions and explanations, the general assistants are fine.
How fast does the market react to news now?
Extremely fast for scheduled, machine-readable releases. Studies put the initial price reaction within about five milliseconds. For messier, unscheduled news that requires interpretation, the reaction can play out over minutes to days, which is where slower sentiment strategies and human judgment still matter.
The bottom line
Quant funds don't have a crystal ball. They have a pipeline that turns words into numbers and numbers into positions, plus the discipline to keep rebuilding it as each edge decays. The two games are speed (latency arbitrage, effectively closed to retail) and signal (sentiment factors, conceptually open to anyone). The academic and vendor data backs up the second game as real but modest: 20-27% sentiment-return correlations, Information Ratios near 4 in good periods, all of it shrinking as the crowd catches up.
What's genuinely new in 2026 is access. The sentiment-and-impact scoring that once lived behind six-figure data contracts now shows up, in simplified form, in consumer tools. You won't out-trade Renaissance with a €24.99 subscription. But you can finally see news the way a model does, as scored, filterable data rather than an anxiety-inducing scroll. That shift in perspective is the real democratization, and it's worth understanding even if you never place a single news-driven trade.
If you want to see it for yourself, spin up a no-card NowNews trial and watch sentiment, impact, and price share one screen for a week.
This article is updated as quant news-trading methods evolve. Last reviewed: June 2026.