How ARIA reads the news: importance, direction and repetition

The problem: almost everything is noise

For a popular asset dozens of pieces come out each day: "price prediction", "top 10 to buy", quote pages, presales. Almost none carries a new fact. ARIA first applies a free pre-filter that drops whatever neither names the asset nor says anything, and only then spends a classification.

Two questions per news item

Each item that passes the filter is classified with two questions, and each answer carries its probability:

• Importance: high, medium or low.
• Direction of the price over one to two days: up, down or neutral.

Keeping the probability, not just the label, tells you when the classifier is unsure.

How it all combines into one reading

Everything classified in 24 hours becomes a single reading per asset, weighted like this:

• Freshness: an item weighs half as much every 12 hours.
• Repetition: what was already reported another day in the same direction weighs half; the opposite direction is a new fact.
• Corroboration: the same story across several outlets counts once, with a small bonus per outlet.
• Social media: weighs half as much as the press, because it carries a lot of noise.

If there is not enough evidence, ARIA gives no opinion. That is a design choice: staying quiet beats making things up.

Is it already priced in?

Good news the market has already digested is not an opportunity. Before proposing a purchase, ARIA checks how much the price has moved since the news: if it already rose more than one standard deviation, it does not enter.

And then, measure

Each classification stores the price at the time. That way you can measure, at 24 and 48 hours, whether "up" readings really rose more than the baseline. Without that measurement, everything above is a hypothesis. More about ARIA: what it is and why it trades on paper.