How AI Detects Market Events
NLP, scoring, delivery — and why direction is left out
10 min read · Last updated August 2026
Artificial intelligence has reshaped how sophisticated traders monitor markets. What once required a team of analysts reading news feeds, watching SEC filings and correlating data across sources can now run as an automated pipeline that processes thousands of events a day and surfaces the ones that look consequential. This guide explains how that technology actually works — including the part of it that did not work.
The Problem It Solves
A person can watch perhaps three information sources attentively. The events that reprice US equities arrive across dozens — SEC EDGAR, a dozen news wires, FDA calendars, earnings feeds, the options tape — at all hours, most of them irrelevant to any given trader. The bottleneck is not analysis; it is noticing.
When a catalyst fires on a ticker outside your coverage, you typically see it minutes after it happened. Automated systems address both halves of that: coverage breadth, by monitoring every source at once, and latency, by delivering a classified, scored event with its exact dispatch time stamped on the alert.
What Gets Monitored
Financial news wires and press releases, FDA PDUFA calendars and approval databases, earnings release feeds, and market data across a broad universe of US tickers, and scheduled macro events. TradeAI News monitors named feeds concurrently, including pre-market and after hours.
How Machines Read Financial News
Natural Language Processing is what lets software read a headline the way an analyst would skim it. In this context, models extract the tickers a story is about, classify the event type — earnings, FDA, M&A, financing, litigation and 35+ others — assess whether the language reads favourably or unfavourably for the company, and estimate how consequential the event appears.
Tone is not direction
That last distinction is worth holding onto: NLP establishes what the text says, not what the price will do. Favourable text and a falling price coexist constantly, because the price already reflected the expectation. Any system that quietly converts a tone reading into a trade direction has made an assumption it probably has not tested.
From Text to Score
Classified events are scored 0–100 on significance — how much this looks like the kind of event that matters. The inputs are the catalyst type and its historical behaviour, how fast the story is propagating, the microstructure around the ticker, and timing relative to the session.
This is a ranking of significance, not a probability of profit, and we do not publish an accuracy figure for it. Tiers: above 82 highest, 68–81 elevated, 55–67 watch, below 55 monitoring only.
Why Direction Is Left Out
This is the part most guides on this topic skip, so here is ours in full.
We ran a pre-registered experiment: given that an event is the kind that produces a large move, is the direction of that move predictable from the data available at detection time? The success criteria — AUC above 0.60, balanced accuracy above 0.57, stable week over week, beating a naive baseline — were fixed before the analysis.
The interpretation is not that our features were poor. It is what efficient-market theory predicts: volatility and magnitude persist and are therefore partly forecastable, while direction is precisely the thing that gets arbitraged away fastest. Establishing that a stock is likely to move sharply is a tractable problem. Establishing which way is, at least with this data, not.
So the product publishes significance and withholds direction. Not as caution — as a finding. The full write-up is here.
Delivery
Detection, classification, scoring and delivery are instrumented end to end — every alert carries its dispatch timestamp. Events reach the dashboard live, Telegram next, and the REST API on Elite for programmatic consumers.
Frequently Asked Questions
What data does an event detection system monitor?
Financial news wires, SEC EDGAR filings, FDA PDUFA calendars and approval databases, earnings release feeds, and scheduled macro events. TradeAI News monitors named feeds concurrently.
How accurate is it?
We publish no accuracy figure for the significance score, because we have not validated one we would defend in public. For direction we published the failure instead: a pre-registered test came in at AUC 0.573 against a required 0.60, unstable week to week, with negative net returns even on the highest-confidence subset. So the system does not publish a direction at all.
What is NLP and how is it used here?
Natural Language Processing is what lets software read financial text the way an analyst skims it: extracting the tickers a story is about, classifying the event type, assessing whether the language reads favourably or unfavourably for the company, and estimating how consequential the event appears. It replaces the reading, not the judgement.
Is the tone assessment a direction?
No. It establishes what the text says, not what the price will do. Favourable text and a falling price coexist constantly, because the price already reflected the expectation. The two are separate measurements and we do not model a link between them.
Can this replace human judgement?
No. It removes the monitoring burden — the part where you try to watch six feeds at once and lose. Evaluating the specific catalyst, sizing the position relative to your account risk, and timing execution all remain entirely yours.
How is this different from technical analysis?
Technical analysis reads price and volume history and is inherently backward-looking. Event detection reads the catalyst itself as it is published. They are complementary approaches, and neither of them tells you direction reliably.
See the detection pipeline in action
Named feeds, dispatch time stamped on every alert. Free plan, no card.
Not financial advice. Trading involves risk.
Last updated September 8, 2026