Infoblock alpha
q-fin/ai-finance IB-arxiv-2607.28496-finbert-nonlinear-vs-linear-f1 measured

On 41,618 news-stock pairs from the FNSPID dataset, FinBERT sentiment features achieve F1=0.576 under nonlinear models but only F1=0.230 under linear models.

Active Evidence state is recorded explicitly and may be revised by a later event.

Observed quantities

news_stock_pairs 41618 · f1_nonlinear 0.576 · f1_linear 0.23

01

Primary source

  1. arxiv · 2607.28496 · v1 · abstract
Title
Beyond Sentiment: Structured Information Extraction from Financial News
Authors
Daohan Zhu, Sitong Ge, Ruofei Wang, Honggu Chen, Yubo Hou, Tao Wan, Zengchang Qin
Version
v1
Check
Support quote found word for word in the arXiv abstract v1, 2026-10-08
Supporting quote from the source
41,618 news--stock pairs from the FNSPID dataset, we find that (i) FinBERT sentiment features exhibit strong predictive power under nonlinear models (F1=0.576) but substantially weaker performance under linear models (F1=0.230)

02

Boundary of this block

No additional public limitation was recorded.

03

Connections

Outgoing

No outgoing public edges.

Incoming

No incoming public edges.