Sentinel tests whether machine-read news can predict markets with a pre-registered live record
IPSUM99 · reddit · 2026-07-28
Sentinel is presented as a live experiment testing whether machine-read news contains usable market-moving information.
- The project measures predictions at the trading-day level rather than treating each headline as an independent sample.
- The author argues the honest sample size is 40 trading days, not tens of thousands of row-level predictions.
- Results are reported against trivial baselines using date-clustered confidence intervals and a threshold-free Information Coefficient.
- The evaluation setup is pre-registered and version-controlled, and all data so far is treated as burned development data.
- The post also says a widely used finance sentiment model produced an inverted aggregate signal during this period, with errors concentrated in its most confident predictions.
The author asks for help reworking the model.
More from Infra
- MinerU turns PDFs and Office docs into LLM-ready Markdown to cut token costs — lxfater · 2026-07-28
- UNIUBI AI’s robot dog reportedly lands a 720° backflip on NVIDIA Orin — Scobleizer · 2026-07-28
- AI spending worries push Samsung and SK Hynix down more than 10% — nordicinst · 2026-07-28
- NVIDIA backs Ilya Sutskever’s SSI with a $5 billion strategic investment — loopuleasa · 2026-07-28
- A sandbox for untrusted AI code held up on escapes, but DNS still leaked data — Ok_Beyond_6313 · 2026-07-28
- Inkling’s open weights are served by multiple vendors, but behavior still varies — acmurthy · 2026-07-28