My research focuses on financial machine learning: evaluating language models, adapting compact models for financial tasks, and testing systematic investment strategies. I emphasise controlled comparisons, realistic execution assumptions and transparent reporting of results and limitations.
I publish as Alina Khaybullina.
ORCID · GitHub · Hugging Face
Financial language models
Momentum, Commentary Tone, and Earnings Surprise in LLM Financial Forecasts
15 October 2026 · arXiv preprint
arXiv:2610.22043 · PDF · Benchmark data · Results explorer · GitHub showcase
This controlled study independently varies price momentum, commentary tone and earnings surprise in synthetic announcements across six Q8 checkpoint conversions. In the primary comparison, tone and momentum each have larger effects on answer scores than earnings surprise in five of six checkpoints. The study measures input reliance under the tested protocol; answer encoding and conversion choices affect the results. It does not establish an internal reasoning mechanism or predictability of realised returns. Public prompts, recorded scores and an interactive explorer accompany the paper.
FinVector-Market-4B: A Controlled Study of LoRA Adaptation for Structured Financial Tasks
6 October 2026 · arXiv preprint
arXiv:2610.08882 · PDF · GitHub · Model adapter · Demo
FinVector adapts Qwen3.5-4B with LoRA and compares the base and adapted models on a frozen 600-example benchmark under matched output contracts. The adapter improves financial question answering, calculator-expression correctness and agreement with synthetic scenario targets. Filing overlap and calculator-target inconsistencies limit generalisation claims; these are task-specific benchmark gains. The public release includes an MLX model adapter, documentation and aggregate evaluation results. Training data and benchmark examples remain private.
Other records: SSRN · Zenodo. Use the arXiv record above for the current paper title.
Systematic investment research
Machine Learning and Dynamic Risk Allocation in Trend Following
2026 · Preprint · Code not publicly released
A gradient-boosted classifier decides whether to retain an eligible allocation or hold cash across eight international equity ETFs. The selected strategy is evaluated from 2016–2025 with annual chronological refits, next-close execution and explicit trading costs. The historical return-risk profile improves on the baseline, but a wide bootstrap interval and weaker later-period returns limit evidence of persistence.
Cross-Asset Shock Diffusion: A Reproducible Test of Residual Underreaction, Shock Coherence, and Trading Economics
2026 · SocArXiv preprint
Paper and DOI — SocArXiv · Code · RePEc · Zenodo archive
The study tests residual underreaction and shock coherence using rolling factor models, next-open execution, transaction costs and randomised-signal comparisons across 24 ETFs. The disclosed baseline provides no robust predictive evidence and becomes economically unattractive after costs. Its contribution is an auditable negative result and a reproducible framework for testing information diffusion, timing integrity and trading economics.
Earlier work
NLP Methods for Information Extraction from Text
2022 · Technical manuscript
A survey of NLP methods for turning unstructured text into usable information, covering probabilistic language models, sequence labelling, vector semantics, word embeddings and text classification. Includes a sentiment-classification exercise using more than 10,000 user reviews.
Last updated: 25 October 2026.