Sentiment Signals in a Fragile Democracy: NLP-Driven Detection of News and Social-Media Effects on Peruvian Banking Asset Volatility

Technical Summary: Sentiment Signals in a Fragile Democracy

Problem Statement
While Natural Language Processing (NLP) has become a standard tool for extracting investor sentiment in advanced economies to inform trading and risk models, its application to banking asset volatility in politically polarized emerging markets remains underexplored. Peru presents a distinct analytical challenge due to its institutional fragility and high "political noise." Between 2021 and 2024, the country experienced extreme governance instability, including five different presidents, multiple impeachment proceedings, a failed coup attempt, and regulatory uncertainty under progressive administrations. Despite the concentration of the Peruvian banking sector (four institutions hold over 83% of deposits) and its sensitivity to political shocks, no prior study has specifically examined the sentiment dynamics of Spanish-language Peruvian news and social media in relation to the Lima Stock Exchange (BVL) banking sub-index, nor the impact of regulatory communications from the Superintendencia de Banca, Seguros y AFP (SBS), the Superintendencia del Mercado de Valores (SMV), and the Central Reserve Bank (BCRP).

Methodology
This paper employs a systematic literature review following PRISMA 2020 guidelines, analyzing publications from 2018–2025 indexed in Scopus, Web of Science, and IEEE Xplore. The review focuses on studies utilizing NLP methods (e.g., FinBERT, BERT, LDA, GARCH variants) applied to financial or banking data with quantitative volatility metrics.

Based on this review, the author proposes a conceptual methodological framework tailored to the Peruvian context, structured in six phases:

1. Data Collection: Aggregating daily data from Twitter/X (Spanish financial/political accounts), Peruvian financial news outlets (El Comercio, Gestión), and official regulatory press releases (SBS, SMV, BCRP).
2. Preprocessing: Applying language detection, tokenization, stopword removal using a custom Peruvian-Spanish financial lexicon, and lemmatization via SpaCy.
3. Sentiment & Topic Modeling: Utilizing FinBERT (fine-tuned on a Spanish financial corpus augmented with Peruvian regulatory language) to classify text as positive, negative, or neutral, and employing Latent Dirichlet Allocation (LDA) to identify latent topics (e.g., macroeconomic, political, regulatory).
4. Volatility Modeling: Fitting an EGARCH(1,1) model to daily BVL banking sub-index returns, incorporating lagged sentiment scores as exogenous covariates.
5. Event Study: Analyzing specific political event windows (e.g., the 2021–2022 Castillo governance crises, the December 2022 coup attempt) to measure abnormal volatility (standardized residuals exceeding ±2.58 standard deviations).
6. Policy Translation: Converting findings into a dashboard framework for regulatory supervision.

Key Contributions
The paper makes three specific contributions:

1. Comparative Evidence Matrix: It constructs a nine-jurisdiction comparative matrix documenting the quantitative performance of NLP-augmented volatility models (including FinBERT, BERT, and GJR-GARCH) against statistical baselines.
2. Contextual Framework: It proposes a conceptual framework specifically tailored to Peru's unique political and regulatory environment, addressing the gap in Spanish-language NLP applications for Peruvian financial text.
3. Proposed Pipeline: It outlines a specific "FinBERT + EGARCH" pipeline designed to integrate Spanish-language preprocessing with BVL banking data, aiming to detect systemic sentiment shocks before they manifest as capital-market distress.

Results and Evidence
The systematic review synthesizes evidence from nine jurisdictions (including the USA, UK, China, South Africa, and Croatia), yielding the following findings:

• Performance Improvement: NLP-augmented models consistently outperform statistical baselines (ARIMA, plain GARCH) in directional accuracy, showing improvements ranging from 14 to 24 percentage points. For instance, FinBERT combined with LSTM outperformed ARIMA in Central European markets, and OPT/BERT models achieved ~72–74% accuracy in US markets compared to 52% for dictionary-based methods.
• Asymmetric Leverage Effect: Negative sentiment consistently exerts a stronger and faster-decaying volatility effect than positive sentiment of equivalent magnitude. This asymmetry is confirmed via GJR-GARCH and EGARCH coefficients in markets like South Africa, China, and the UK.
• Predictive Power of Political Uncertainty: Social media-based political uncertainty measures (e.g., Twitter-based Market Uncertainty) significantly predict volatility in 18 of 20 international equity markets.
• Peru-Specific Projections: Based on international analogues, the paper projects that the proposed FinBERT + EGARCH model for the BVL banking sub-index could achieve a directional accuracy of approximately 82.1%. This represents a 41% improvement over an ARIMA baseline (58.2%) and a 10.5% improvement over standalone FinBERT.
• Event Impact: The review cites evidence that the December 2022 attempted coup in Peru generated intraday return variability of 4.26 standard deviations above the baseline for the BVL General Index, with banking stocks being particularly exposed.

Significance and Claims
The paper argues that NLP-based sentiment monitoring offers a viable, cost-effective approach for early-warning systems in Peruvian banking supervision. It posits that regulatory bodies (SBS and SMV) would benefit from deploying FinBERT-augmented EGARCH dashboards to detect sentiment shocks 48–72 hours before they translate into observable volatility spikes. This is particularly critical in a context where regulatory bodies themselves are sources of policy uncertainty.

The authors explicitly state that this study is conceptual-comparative in design. While the evidence from other jurisdictions strongly suggests the efficacy of the proposed framework, the paper acknowledges that empirical validation using actual BVL banking sub-index and Twitter corpus data for the 2021–2024 period remains the essential next step. The proposed performance metrics are projections derived from linear interpolation of international data and require future empirical testing with pre-registered protocols.