How Event Detection Works
NLP, classification, delivery — and why direction is left out
10 min read · Last updated August 2026
Software has changed how 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 reads every item from many feeds and classifies them by type. 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 many sources — SEC EDGAR, news wires, FDA calendars, earnings feeds — at all hours, most of them irrelevant to any given trader. The bottleneck is not analysis; it is noticing.
An event on a ticker outside your coverage is easy to miss. Automated systems help with coverage breadth, by reading every source at once, and with consistency, by classifying every event the same way.
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 more (TradeAI News observed 40 distinct event types in the last 30 days) — and assess whether the language reads favourably or unfavourably for the company. TradeAI News itself does this with keyword and phrase rules rather than a language model, and reads the text’s tone with FinBERT, a financial-language model.
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 Classified Event
Classified events get an internal value that orders which are shown first. The value itself is never displayed, and it has not been validated as a predictor.
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 an event, is the direction of the price move that follows predictable from the data available at detection time? The success criteria were written down before the analysis.
The interpretation is not that our features were poor. It is what efficient-market theory predicts: volatility tends to cluster, while direction is what every participant competes to price first. With this data, direction was not predictable.
So the platform publishes measured structure and withholds direction. Not as caution — as a finding. The full write-up is here.
Delivery
Classified events appear on the dashboard, labelled by type, next to the same ticker’s other data. For programmatic use they are also available through the REST API.
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 internal value that orders events, because we have not validated one we would defend in public. For direction we published the failure instead: a pre-registered test failed every one of its criteria (the protocol and the results are on our page “Why we don’t predict direction”). 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, and assessing whether the language reads favourably or unfavourably for the company. It replaces the reading, not the judgement. TradeAI News itself does not use a language model to classify events today: it uses keyword and phrase rules (a Bayesian keyword pre-filter plus fixed phrase rules), and reads the text’s tone with FinBERT, a financial-language model.
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 event, 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 event itself, from the source. They are complementary approaches, and neither of them tells you direction reliably.
See the whole platform
Gamma, insiders, volume, short interest, events and your journal. Free plan, no card.
Not financial advice. Trading involves risk.
Last updated October 14, 2026