Claim Missing Document
Check
Articles

Found 1 Documents
Search

Hybrid Sentiment Analysis Berbasis LLM sebagai Data Alternatif untuk Mendukung Nowcasting Inflasi Provinsi Jawa Tengah Nuki Pujiani Yosephine; Zamrud Mahfur Abdillah; Fayza Nayla Riyana Putri; Marita Prasetyani
Seminar Nasional Official Statistics Vol 2026 No 1 (2026): Seminar Nasional Official Statistics 2026
Publisher : Politeknik Statistika STIS

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.34123/semnasoffstat.v2026i1.2860

Abstract

Inflation is a key macroeconomic indicator for economic policy. This study evaluates online news as alternative data for supporting inflation nowcasting in Central Java through weighted fusion of a lexicon score and few-shot Large Language Model classification (Claude Sonnet). Of 147 raw Google News RSS records, 87 fell within April 2024-June 2026, and 84 articles from 27 outlets passed topical and geographic verification. A 0.0-1.0 weight simulation selected 90% lexicon and 10% LLM by maximizing the absolute contemporaneous correlation. The hybrid score had a moderate inverse correlation with month-to-month inflation (r=-0.545; p=0.067; n=12), stronger than the lexicon-only (r=-0.517) and LLM-only (r=- 0.190) scores, but it was not statistically significant. Calendar-based lag exploration from zero to five months found a significant two-month association only for the LLM-only score (r=-0.864; p=0.001; Holm-adjusted p=0.022; n=10); the hybrid score was not significant at any lag. Thus, fusion strengthened the contemporaneous association in this sample but did not establish predictive superiority, requiring validation on a longer time series.