Trading strategies·event

News trading

◆still alpha

Reviewed 4 June 2026. As of 2026: a real edge still exists for those who can run it well.

React to information releases via machine-readable feeds: structured data, low-latency wires, and increasingly NLP. The edge is speed and accuracy of interpretation, not having the news first.

The idea

News trading annotated diagramfigure

React to information releases via machine-readable feeds: structured data, low-latency wires, and increasingly NLP. The edge is speed and accuracy of interpretation, not having the news first.

Reference figure. This concept is explained in prose and diagram; the interactive widgets live on the flagship pages it links to under Where this fits.

What is machine-readable news?

Machine-readable news is news delivered as structured data, not prose: each item tagged with the entities it concerns, the event type, a sentiment/relevance score, a novelty flag and a timestamp, so a machine can act on it without reading English. Vendors (Bloomberg, Reuters/Refinitiv, Dow Jones Newswires) "elementize" headlines specifically so algorithms can trade them in microseconds. This is the data substrate of directional event trading's reaction leg.

Intuition first. A trading system cannot read a paragraph in time, so news vendors pre-process headlines into fields: which company, what happened (earnings/M&A/guidance/litigation), how positive, is this new or a repeat, exact timestamp to the microsecond. The algorithm consumes the fields, not the sentence. The major feeds in 2026 are the Dow Jones Newswires / DJ elementized feeds, Bloomberg (news plus the BLPAPI / event-driven feeds), and Reuters/Refinitiv (LSEG) News Analytics, each delivering low-latency, tagged items over a co-located connection, built explicitly for algorithmic consumption. The exact fields and latencies change, so the vendor's own product spec is the source to cite.

An elementized item is a tuple, not a sentence: the resolved entity, the event class, a signed sentiment/relevance score, a novelty flag and a microsecond timestamp. The algorithm reads the tuple and never parses English on the hot path.

\text{item} = (\,\text{entity},\; \text{event type},\; \text{sentiment},\; \text{novelty},\; \text{confidence},\; t\,)

The split that matters runs through the whole page. Structured numeric releases (an earnings number, an economic figure) are trivial to parse and the surprise is arithmetic; this is the scheduled-event game, won on network latency. Unstructured headlines (a takeover rumour, a regulatory action, a CEO quote) require interpretation; this is where NLP and LLMs earn their place, and where 2026's frontier sits.

How is news turned into a trade? (the pipeline)

A news-trading system runs a pipeline: ingest the feed, parse/classify the item (entity, event type, sentiment, novelty), score confidence, map to a trade decision, then submit, all under a tight latency budget. Each stage adds latency; the binding constraints are interpretation (the hard, slow step for unstructured news) and false-signal control (gating on confidence so a wrong read does not trade).

Ingest: receive the item on the lowest-latency channel (a co-located feed handler) and deduplicate against items already seen, since wires repeat and cross-publish. Parse/classify: for structured items, read the fields; for unstructured items, run NLP: entity resolution (which listed company), event classification (M&A? guidance? litigation?), sentiment/direction, and novelty (is this genuinely new information or a restatement?). Novelty is critical: the market already priced yesterday's rumour. Confidence gate: score how sure the classification is and fire only above a threshold, deferring or skipping ambiguous items: the single most important defence against false signals. Decision and submit: map the entity, direction, magnitude and confidence to a sized order and submit, managing the thin, fast-moving book at news time (realistic fills; see backtesting).

Total time-to-trade is the sum of the pipeline stages. For structured items the network terms dominate (microseconds); for unstructured items the parse/interpret term dominates (milliseconds of model inference), which is exactly the LLM trade-off below.

t_{\text{trade}} = t_{\text{ingest}} + t_{\text{parse}} + t_{\text{decide}} + t_{\text{submit}}, \qquad t_{\text{parse}} \gg \text{rest (unstructured)}

The latency budget is the design constraint: for structured items, network latency dominates and the field competes in microseconds; for unstructured items, interpretation latency (model inference) often dominates network latency. Only fire on novel, market-moving items above the confidence threshold; speed without a confidence gate is negative edge.

What do NLP and LLMs actually change in 2026?

NLP/LLMs genuinely improve the interpretation half of news trading: reading novel, unstructured headlines correctly (entity, event, direction, novelty) better than keyword/lexicon methods. But they add inference latency (milliseconds to act, versus microseconds for structured feeds), they hallucinate (false signals), and they commoditise the easy reads. They shift, not remove, the bottleneck, and they do not beat a co-located numeric feed on speed.

What genuinely improves: classifying unstructured news, disambiguating "Company X explores sale" (takeover, bullish) from "Company X faces fire-sale pressure" (distress, bearish), resolving which entity is meant, judging novelty against context. Older lexicon/keyword sentiment (e.g. the Loughran–McDonald financial dictionaries, 2011) is brittle on phrasing; transformer/LLM classifiers read context. This is a real lift on the interpretation axis. What it does NOT change: the speed axis for structured signals. An LLM cannot out-race a co-located system reading a tagged numeric release: inference takes milliseconds; the structured-feed firm is done in microseconds. AI helps where the bottleneck is understanding, not where it is network distance.

The new costs AI introduces are concrete. Inference latency: running a model per headline costs milliseconds (an eternity in HFT) so firms distil/quantise models, pre-filter with cheap classifiers, or reserve the LLM for ambiguous items only. Hallucination / false signals: an LLM can confidently misclassify, and acting on it is a directional loss, so the confidence gate and human-audited guardrails matter more, not less. Crowding: once everyone has capable LLMs, the easy interpretive edges are competed away, and the lift accrues to whoever has better data, faster inference and tighter gating, not to "having an LLM".

AI buys accuracy at the cost of latency. It is worth firing the model only when the extra correctness it brings outweighs the move lost to its inference time, which is why LLMs are reserved for ambiguous, unstructured items, not tagged numeric releases.

\Delta\mathbb{E}[\pi] \;=\; \underbrace{\Delta p \cdot \Delta P}_{\text{accuracy gain}} \;-\; \underbrace{\text{decay}(t_{\text{infer}})}_{\text{move lost to latency}}

The honest verdict: AI moves the news-trading edge toward correct interpretation of hard, novel news and away from raw speed on easy, structured news, but it is an arms race, not a free lunch. See machine learning in HFT and what AI changes for HFT.

Where is the edge in 2026, honestly?

Thin and specific. Structured numeric news is arbitraged in microseconds by co-located firms, leaving no edge for the latency-disadvantaged. The surviving retail-accessible edge is interpretation of unstructured, novel news in less-efficient venues (small caps, crypto, prediction markets) where the field is slower and the right read beats the fast-but-wrong. False-signal control, not speed, is the differentiator there.

No edge: trying to out-speed institutional co-located feeds on tagged numeric releases, where you will lose the race. Possible edge: correctly interpreting novel, unstructured news faster than the slower field in venues where milliseconds are not decisive, such as small caps, crypto, and prediction markets (where a headline can resolve or reprice a contract and the book is thin). Here being correct matters more than being microseconds-fast.

The trap: backtested news strategies routinely overstate the edge because they (a) use revised/clean historical news that did not look that clean in real time, (b) assume fills the thin event-time book would not give, and (c) ignore the false-signal rate that only shows up live. Model the real-time feed, realistic fills, and a true false-positive rate (see backtesting & simulation). For the brand-level verdict see is HFT still profitable in 2026.

Worked example

A synthetic news-trade decision, as of 2026, with illustrative numbers. At 11:43:17.000 a wire prints: "Company X to acquire Company Y for $4.2bn cash, 35% premium." Ingest + dedup: $+1.2$ ms on a co-located handler; not a repeat of an earlier rumour, so novelty is high. Classify: entity $=$ X (acquirer) and Y (target); event $=$ M&A confirmed; direction for Y strongly positive (premium $\approx +35\%$ ), for X mildly negative; confidence $= 0.92$ . The structured/keyword path takes $\sim 2$ ms; the LLM-assisted path (used because the phrasing was non-template) takes $\sim 18$ ms of inference.

Gate: confidence $0.92 \gt 0.85$ threshold, so trade: buy Y, optionally trim X. Submit: $+0.4$ ms to the matching engine. Total $\approx 4$ ms (structured) or $\approx 20$ ms (LLM-assisted). Outcome: Y reprices $+30\%$ toward the offer over the next seconds. The 4 ms path captures most of the first move; the 20 ms LLM path captures less of the jump but read a non-template headline a keyword system might have missed entirely: the classic speed-versus-correctness trade-off.

The false-signal counter-case: a near-identical-looking later headline, "Company X denies it will acquire Company Y", must be classified as negative/no-event, confidence 0.95 on "denial". A keyword system that matched "acquire" and bought would take a loss. This single example is the difference between speed and correct speed.

\text{``acquire''} \;\not\Rightarrow\; \text{buy}: \quad \text{``X \textbf{denies} acquire Y''} \Rightarrow \text{sentiment} \lt 0

The live event/duration toy for this family (IX-DURATION) lives on irregular time; this page is diagram-only. Numbers are synthetic and illustrative. Real feed latencies, classification accuracy, inference times and false-signal rates must be measured per feed/model and dated, and the vendor's spec is the source for feed fields and latency. Educational only, not investment advice.

Where this fits

↑ Up · building block of Directional event trading ↑ Up · building block of Trading strategies ↔ Across · composes with Event arbitrage ↔ Across · composes with Machine learning in HFT ↔ Across · composes with Irregular time & point processes → Apply · makes money in Prediction markets (Polymarket) → Apply · makes money in Crypto market making ⤓ Build / Buy · tool needed Datasets & tools

Common questions

Can a machine read the news?

News trading reacts to information releases; in HFT it relies on machine-readable feeds: structured economic data, low-latency news wires, and increasingly NLP/LLM parsing of unstructured text. The edge is speed and accuracy of interpretation, not having the news first. In 2026, language models genuinely help off the hot path (classifying, summarising), but the microsecond reaction itself still runs on pre-computed, classical logic.