From Noisy News Sentiment Scores to Interpretable Temporal Dynamics: A Bayesian State-Space Model

TL;DR

This paper introduces a Bayesian state-space model that transforms noisy weekly news sentiment scores into a smoothed, interpretable time series with explicit uncertainty, accounting for varying news coverage over time and categories.

Contribution

It presents a novel Bayesian framework that explicitly models observation uncertainty based on news coverage, improving sentiment analysis from aggregated news data.

Findings

Latent sentiment dynamics are similar across categories.

Observation noise varies significantly with news coverage.

The model effectively captures uncertainty related to data availability.

Abstract

Text-based sentiment indicators are widely used to monitor public and market mood, but weekly sentiment series are noisy by construction. A main reason is that the amount of relevant news changes over time and across categories. As a result, some weekly averages are based on many articles, while others rely on only a few. Existing approaches do not explicitly account for changes in data availability when measuring uncertainty. We present a Bayesian state-space framework that turns aggregated news sentiment into a smoothed time series with uncertainty. The model treats each weekly sentiment value as a noisy measurement of an underlying sentiment process, with observation uncertainty scaled by the effective information weight $n_{tj}$: when coverage is high, latent sentiment is anchored more strongly to the observed aggregate; when coverage is low, inference relies more on the latent dynamics and uncertainty increases. Using news data grouped into multiple categories, we find broadly similar latent dynamics across categories, while larger differences appear in observation noise. The framework is designed for descriptive monitoring and can be extended to other text sources where information availability varies over time.