Key insights
- A retired software developer built an AI-powered news classifier to identify predictive signals in AI-related financial news based on story framing (growth, risk, etc.) rather than simple sentiment. Early observations suggest shifts in story type precede sentiment changes. A formal test is underway, with results expected in June. The shift in AI coverage from capability to commercialization is also noted.
I'm retired and have been writing software since the early 80s. A few months ago I got curious about whether AI-sector financial news has any predictive structure beyond simple bullish/bearish sentiment.
Not sentiment in the usual sense exactly more like what kind of story journalists are collectively telling. Earnings story, growth story, regulatory-risk story, market-correction story, etc.
So I built a pipeline that pulls in a few hundred AI-related articles every morning, runs sentiment scoring, classifies them into eight story types, then aggregates everything across six buckets: semis, hyperscalers, enterprise software, GPU cloud, datacenter infrastructure, and pure-play AI companies.
One thing I didn't expect: the mix of story types sometimes shifts before the aggregate sentiment does. I first noticed it while debugging the classifier in March. There were a couple of rotations where coverage moved from growth/earnings framing toward risk framing about a day or two before the overall sentiment scores rolled over.
No idea whether that translates into prices yet. Could easily be noise.
I pre-registered the test before looking at returns data because I know how easy it is to fool yourself with this kind of thing. Earliest analysis date is June 1 once I have ~60 trading days.
The other thing that's been interesting is how much AI coverage changed in April. Earlier coverage was dominated by capability narratives; now it's much more deployment/revenue/commercialization focused, and that shift has persisted longer than I expected.
Dashboard is at https://knowentry.com/canary-dashboard/ if anyone wants to take a look at it. No signup/paywall/cookies it's a personal site. Mostly curious whether anyone here has worked on similar news-framing problems or knows relevant literature I should read.