Financial News Summarization: Can Extractive Methods Still Offer a True Alternative to LLMs?

In an era dominated by generative AI, a new study asks whether simpler extractive techniques can still hold their ground in financial news summarization—and what that means for efficiency, accuracy, and real‑world deployment.

</div>

1. The Study at a Glance

Published December 9, 2025, by Reche, Linhares‑Pontes, and Torres‑Moreno, the paper investigates whether extractive summarization methods remain viable alternatives to large language models (LLMs) for financial news summarization . Using the FinLLMs Challenge dataset—8,000 training and 2,000 test samples of short financial news with single‑sentence summaries—the study benchmarks methods ranging from heuristic extractive baselines to fine‑tuned abstractive LLMs .

2. Extractive vs. Abstractive: What the Numbers Reveal

On the test set, simple extractive methods like Lead‑1 and MatchSum achieve respectable ROUGE‑1 scores of ~0.247 and ~0.241, respectively, with BERTScores around 0.588 and 0.583 . In contrast, most off‑the‑shelf abstractive models—T5‑small, Pegasus‑xsum, BART‑large‑xsum—underperform relative to Lead‑1, revealing the limitations of generic LLMs in financial contexts . However, a fine‑tuned Mistral‑7B (FT‑Mistral‑7B‑Instruct‑v0.3) achieves a striking ROUGE‑1 of 0.514 and BERTScore of 0.728—far surpassing extractive baselines .