Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.39636

Information Flow and Market Risk: Transformer-Based Evidence on Sentiment and Volatility Dynamics

Szymon Lis·Robert Ślepaczuk·Paweł Sakowski
View original source ↗Read source document ↗Export metadata ↓

Abstract

This study examines whether transformer-based financial news sentiment improves equity volatility forecasts beyond conventional market-based predictors. Using data for eleven US stocks, we compare volatility persistence models with specifications incorporating the VIX, AAII investor sentiment, and news sentiment extracted using ModernFinBERT. The forecasting target is annualized volatility constructed from future averages of trailing 252-session Yang-Zhang estimates, evaluated at horizons of 1, 5, and 22 trading days. Models are assessed through walk-forward testing with expanding and rolling estimation windows, fixed initial hyperparameter selection, and statistical comparisons accounting for multiple testing and serial dependence. Results show that a naive persistence benchmark achieves the lowest average QLIKE loss at all horizons. Although HAR-VIX produces the largest number of individual forecasting wins under expanding windows, it does not achieve the lowest mean loss. Adding text-based sentiment improves HAR forecasts in only two of 33 expanding-window comparisons, and neither improvement remains statistically significant under longer HAC bandwidths. A blinded audit of 2,700 news texts further characterizes the agreement and confidence patterns of ModernFinBERT classifications. The findings do not establish a general incremental forecasting advantage from transformer-based news sentiment. They highlight the importance of strong benchmarks, temporal alignment, robust inference, and careful evaluation of text-derived predictors.

Explore related subjects

q-fin.RM

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Szymon Lis, Robert Ślepaczuk, Paweł Sakowski. 2026-10-29. Information Flow and Market Risk: Transformer-Based Evidence on Sentiment and Volatility Dynamics. https://arxiv.org/abs/2610.39636

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Decorrelating the Advice Layer: A Calibrated Fragility Metric and Tested Mitigations for Synchronized Liquidation in Agentic Wealth Management

Conversational AI agents are turning wealth management into a continuous loop of market analysis, allocation, risk checks and execution, run for millions of clients on a handful of foundation models. We build a multi-asset agent-based model of this layer, calibrated to measured model concentration, cross-model correlation, retail turnover and 33 years of real equity returns, and ask when the layer amplifies market declines and what prevents it. The effective number of independent decision-makers is bounded however many clients the layer serves: about 2.0 at today's consumer concentration. This monoculture of views drives crowding in individual securities and sets the blast radius of a compromised model provider, but it does not drive market-wide liquidation. Index tail losses rise with the square root of agentic adoption, are 1.47-1.90 times larger under an agentic than under a periodic robo-advice architecture, and do not rise with model concentration; synchronized automation (risk overlays, drawdown cuts and attention-driven execution) drives them. A fragility index computed from aggregate flow, public returns and an impact estimate forecasts tail liquidation out of sample (hold-out R^2 of 0.56) far better than concentration or correlation measures. Decorrelating the automation, not the models, works: dispersing overlay parameters, staging execution and a participation breaker together cut tail liquidation by 63% at no measurable cost to client welfare, whereas provider diversification has no effect. Because the statistics required are sums, they can be computed across platforms under secure aggregation with differential privacy at negligible cost for realistic client populations. In historical replays, an agentic layer holding 10% of the market would have deepened the 2020 crash by about 858 basis points.

q-fin.RM↗
arXiv2026

LLM-Generated Counterfactual Stress Scenarios for Portfolio Risk Simulation via Hybrid Prompt-RAG Pipeline

We develop a transparent, auditable large language model (LLM) pipeline for macro-financial stress testing. Structured prompting, with optional retrieval of country fundamentals and news, emits G7 macroeconomic scenarios (GDP growth, inflation, policy rate), which are converted into deviations from an IMF World Economic Outlook baseline and translated into portfolio losses through a three-channel PCA factor model. The macro-to-factor mapping is stipulated rather than estimated, so every tail-risk multiple is conditional on stated assumptions. Across 2,420 accepted scenarios from two model families and three runs, plus an unanchored-severity run and eight risk-layer ablations, we find three things. First, severity is inherited from scenario design, not generated by the model: prompt variant dominates every variance decomposition, removing the prompted severity bands makes GPT-5.6-luna's recession 1.6 pp milder, apparent country heterogeneity largely reflects the institutional baseline, and retrieval and news matter only where the prompt leaves magnitude free. Second, numerical severity and narrative severity are only weakly coupled. Third, GPT-5.6-luna follows its own arithmetic but anchors on prompted bands, whereas Llama-3.1-8B is four to seven times more dispersed and contradicts its own stated changes in 90% of cases. A repeat run differs at the scenario level yet agrees on aggregate multiples to within 0.007 times. LLM scenario generators are best used for narrative and sector elicitation around a prescribed severity path, not as autonomous calibrators of stress magnitude.

q-fin.RM↗
arXiv2026

Distributions or Quantiles? A Convex Variational Theory of Worst-Case Risk

Worst-case risk evaluation is central to finance and operations research, yet existing analytical results remain fragmented and largely confined to ambiguity sets with specific structures. Using spectral risk measures as the underlying class, we develop a convex variational theory that unifies these analyses and extends them to richer ambiguity models. The theory rests on a representation principle: use the quantile or distribution formulation that preserves convexity of the full ambiguity set. In both representations, the resulting optimality conditions provide a systematic route to identify the worst-case law and determine its risk value. This approach recovers classical central-moment bounds and yields closed-form or semi-closed-form solutions for practically motivated models involving Wasserstein ambiguity with tail stress, two-sided expected shortfall (ES) bands, asymmetric partial moments, geometric-moment bounds, and exponential- or Gaussian-reference Kullback--Leibler (KL) neighborhoods. Beyond these analytical results, the same theory yields projection-based algorithms with fixed-grid convergence under the stated assumptions. In a high-dimensional portfolio application, we establish an exact projection identity for mean-anchored joint KL ambiguity under a Gaussian reference. The resulting standardized worst-case ES coefficient applies to every portfolio, reducing robust portfolio selection to an exact second-order cone program after a single calibration.

q-fin.RM↗