Bayesian calibration for banking financial predictions using large language models: A study on the Vietnamese market
Multidisciplinary Science Journal · 13.10.2026 · 9 (4) · 2027251
DOI: 10.31893/multiscience.2027251 ↗Kurz gefasst
Fachartikel in der Zeitschrift Multidisciplinary Science Journal, online erschienen am 13.10.2026. Verfasst von Nguyen Van Dat. Die Arbeit nennt 22 Literaturangaben.
Abstract (Originalsprache)
Financial predictions from large language models (LLMs) typically lack uncertainty quantification, rendering point estimates unreliable for investment decision-making. This study proposes a Bayesian calibration framework to improve the reliability of LLM predictions for the Vietnamese banking sector. The methodology comprises three models: (1) beta-binomial for direction calibration, (2) Bayesian linear regression for value calibration, and (3) a hierarchical model for five commercial banks (VCB, BID, CTG, MBB, and TCB). The data include financial reports and stock prices from 2019–2023. The results show that the hierarchical model (Eq. 3) achieved the lowest expected calibration error (ECE = 0.22), a 45% improvement over the raw LLM predictions (ECE = 0.40) on a five-bank-year (N=5) test partition, with continuous ranked probability score (CRPS) = 0.113. Priors follow weakly informative specifications: α ∼ N(0, 10), β ∼ N(1, 5) centered at the unit slope, σ ∼ HalfNormal(1), with hierarchical hyperpriors τ_α, τ_β ∼ HalfNormal(5). Posterior inference uses the no-U-turn sampler with 2,000 draws across 4 chains and 1,000 tuning steps; convergence is monitored at R̂ < 1.01 and ESS > 400. The hierarchical model borrows statistical strength across banks via partial pooling, a property well suited to small-sample emerging-market settings where each institution yields only five annual observations. Practical implications: Investors receive calibrated probability ranges rather than overconfident point forecasts, enabling defensible portfolio allocation under uncertainty. Methodological implications: Partial-pooling hierarchies can be routinely applied to sparse financial time series. Theoretical implications: To our knowledge, this is the first empirical evidence that Bayesian calibration materially improves LLM forecast reliability in an emerging market. We transparently report Eq. 2 convergence failure (R̂ = 4.1) as a sample-size caution. This research contributes to the first Bayesian calibration framework for LLM financial predictions in Vietnam, providing credible intervals instead of point estimates. Because the GPT-4o-mini knowledge cutoff overlaps the 2023 test year, the reported 45% gain should be read as an optimistic upper bound, pending evaluation on a fully out-of-cutoff horizon.
Zitieren
APA
Dat, N. V. (2026). Bayesian calibration for banking financial predictions using large language models: A study on the Vietnamese market. Multidisciplinary Science Journal, 9(4), 2027251. https://doi.org/10.31893/multiscience.2027251
Vancouver
Dat NV. Bayesian calibration for banking financial predictions using large language models: A study on the Vietnamese market. Multidisciplinary Science Journal. 2026;9(4):2027251. doi:10.31893/multiscience.2027251
BibTeX
@article{Dat2026,
author = {Dat, Nguyen Van},
title = {Bayesian calibration for banking financial predictions using large language models: A study on the Vietnamese market},
journal = {Multidisciplinary Science Journal},
year = {2026},
volume = {9},
number = {4},
pages = {2027251},
doi = {10.31893/multiscience.2027251}
}
EUVIMED