Can news become a trading signal? Part 1: the methodology

0:00After the Cambridge Analytica revelations,

0:02Facebook stock fell nearly seven

0:03percent in one day.

0:05Mark Zuckerberg's net worth dropped

0:06about five billion dollars that

0:07Monday. We know news can

0:09move stocks. But by how

0:10much, how reliably, and could

0:13a trading strategy use it?

0:14Here's how we're testing it.

0:16First, we needed a way

0:17to measure the news.

0:18Our archive held more than

0:19a hundred million headline and

0:21target matches. Sending every one

0:22to a language model would

0:23mean paying for, and waiting

0:25on, millions of separate judgments.

0:27Even cheap models can disagree

0:29about what counts as good

0:29news. So we used language

0:31models to label a much

0:32smaller training set, then trained

0:33a classifier to do the

0:34repeated work. It answers two

0:36questions: is this story really

0:38about this company, and is

0:39the news positive, negative, or

0:41neutral for it?

0:42The first approach was quick

0:43and cheap: Qwen text embeddings,

0:45with a small neural network

0:47on top. Think of an

0:48embedding as a numerical summary

0:49of a headline.

0:50Cache it once, and testing

0:52another small classifier costs almost

0:53nothing. But it kept missing

0:55the relationship. A headline about

0:57Extreme Networks winning business against

0:58H P E was called

0:59positive for both companies.

1:01And even on its own

1:02training examples, our retained company

1:04model scored just seventy-three point

1:05nine percent. More training could

1:07not put the missing meaning

1:08back into those frozen summaries.

1:10We switched to DeBERTa, a

1:11smaller model trained to read

1:13the actual headline together with

1:14the company we're asking about.

1:15Mark the target, read the

1:17words together, and classify that

1:18relationship. The same headline can

1:20now be positive for Extreme

1:21Networks and negative for H

1:23P E. The training data

1:24mattered too. We removed bad

1:26company matches, added fresh news,

1:28and gave the labeler a

1:29consistent investor rubric.

1:31A sports medal isn't silver-industry

1:33news. A routine product launch

1:35isn't automatically a positive investment

1:36signal. On unseen stories, the

1:38selected model agreed with our

1:39Haiku labels on eighty-five point

1:41seven percent of company examples

1:42and eighty-seven point two percent

1:44of industry examples.

1:45That's agreement with the labeler.

1:47Then comes scale: score the

1:48archive once, store the results,

1:50and process only new matches

1:52each day. The initial backfill

1:53budget was five to fifteen

1:55dollars of GPU time.

1:56We'll show that estimate beside

1:57the measured run, because moving

1:59and preparing the data takes

2:00time too. Now those scores

2:01can become strategy inputs.

2:03We still have to respect

2:04when the news was available,

2:05handle missing coverage, and test

2:07whether any relationship survives on

2:09later dates. Classification is the

2:10first measurement step.

2:12The full methodology is linked

2:13in the article on NexusTrade.

2:15Follow for part two, where

2:17I share the full results.