Skip to Content
SignalWise

How reasoning is generated

Search, write, verify, translate. The method behind each explanation.

Collect the context

When an alert fires, SignalWise collects its context: the asset, its type and exchange, the alert type, the direction and the time.

Search vetted news sources

SignalWise searches a curated list of trusted financial news sources in several languages. Only news published in a short window before the alert is used: from a few hours for sharp moves to a few days for 52-week highs and lows.

For assets listed in Tokyo, the search also runs in Japanese on Japanese financial media. Price quote pages, listing pages and crowd-sourced forecast sites are excluded.

Identify the drivers

An AI model reads the relevant articles and identifies the most likely drivers of the move, in this order of priority:

  1. News about the company itself (earnings, contracts, corporate actions)
  2. News about its sector or industry
  3. Macro or geopolitical news, only when the link to the asset is clear

Verify every reason

Each reason must pass automatic checks. A reason that fails any check is removed:

CheckQuestion it answers
Fact dateDid the fact happen inside the alert's time window? Undated facts are removed.
EntityIs it about the right company, pair or commodity, and not a company with a similar name?
DirectionDoes it explain a move in the same direction as the alert?
QualityIs it specific and sourced, not generic or invented?
RelevancyHow strongly does it explain the move (score 1 to 5)?

Translate

Reasons are written in English, then translated into your languages with finance-specific rules: translate the meaning rather than word by word, keep the sentiment, keep tickers and names unchanged. Japanese and Korean have dedicated style rules (for example です・ます style and official Japanese company names).

Deliver

The remaining reasons (up to 3) are ordered by relevancy and attached to the alert.

Quality control

  • Every change to the method is tested against a reference set of alerts before release.
  • Every generation keeps a step-by-step trace, so any reason can be investigated.
  • Native speakers review language quality for key languages.
  • Back-tests on your own universe can be run on request.

Reasoning is generated once per event

The same market event gets the same reasoning for every client covering that asset. It is not personalized per user.

On this page