← Back to VOLUME 15, ISSUE 11, NOVEMBER 2026
This work is licensed under a Creative Commons Attribution 4.0 International License.
FinNews: Detecting Second-Order Supply-Chain Shocks Using Knowledge-Enriched Cross-Encoders
Ramya Bharathy, Abhilash D, Kamalesh Kumar K*, Deepashri
π 5 viewsπ₯ 2 downloads
Abstract: Investor news filters usually match company names, tickers or keywords, or compare embeddings of a query and a headline that were computed separately. These methods handle company announcements well but miss a different kind of risk: a problem at a supplier, a raw-material source or a transport route that affects a company without the headline ever naming it. We call these second-order shocks. FinNews detects them by describing each monitored company with a three-part text profile (identity, sector and upstream inputs) and scoring every headlineβprofile pair with a cross-encoder, which reads both texts together instead of relying on name overlap. The scorer is a DeBERTa-v3-small model initialized from a natural language inference (NLI) checkpoint and fine-tuned on 66,606 pairs covering 55 companies in 10 sectors, split chronologically into training, validation and holdout sets. On the 6,661-pair holdout set at a threshold of 0.40, the model reaches 74.06% accuracy, 71.67% precision and 79.16% recall; it retains 77.78% of second-order shocks and rejects 69.01% of irrelevant headlines. A 12-layer MiniLM passage-ranking baseline, scored on an earlier holdout without enriched profiles, retains 47.60% of second-order shocks. Because the backbone and the profile format changed together between these runs, the gain cannot yet be assigned to either. On a small supplementary live set collected on 10β12
shocks (83.33%). We also report an evaluation error in which a second sigmoid applied to already-normalized scores makes every pair pass the threshold, and we describe a CPU-only deployment that scores a pair in 46β66 ms.
Keywords: financial news filtering, supply-chain risk, second-order events, cross-encoder, DeBERTa-v3, natural language inference
shocks (83.33%). We also report an evaluation error in which a second sigmoid applied to already-normalized scores makes every pair pass the threshold, and we describe a CPU-only deployment that scores a pair in 46β66 ms.
Keywords: financial news filtering, supply-chain risk, second-order events, cross-encoder, DeBERTa-v3, natural language inference
How to Cite:
[1] Ramya Bharathy, Abhilash D, Kamalesh Kumar K*, Deepashri, βFinNews: Detecting Second-Order Supply-Chain Shocks Using Knowledge-Enriched Cross-Encoders,β International Journal of Advanced Research in Computer and Communication Engineering (IJARCCE), DOI: 10.17148/IJARCCE.2026.151164
