Reasoning on Time-Series for Financial Technical Analysis
Kelvin J.L. Koa, Jan Chen, Yunshan Ma, Huanhuan Zheng, Tat-Seng Chua
TL;DR
This paper introduces Verbal Technical Analysis (VTA), a novel framework combining verbal and latent reasoning to produce accurate, interpretable stock time-series forecasts by converting price data into textual annotations and optimizing reasoning traces.
Contribution
VTA is the first framework to integrate verbal reasoning with time-series forecasting, enabling interpretable and accurate stock predictions across multiple markets.
Findings
VTA achieves state-of-the-art forecasting accuracy.
Reasoning traces are validated as meaningful by industry experts.
The framework generalizes across different international stock markets.
Abstract
While Large Language Models have been used to produce interpretable stock forecasts, they mainly focus on analyzing textual reports but not historical price data, also known as Technical Analysis. This task is challenging as it switches between domains: the stock price inputs and outputs lie in the time-series domain, while the reasoning step should be in natural language. In this work, we introduce Verbal Technical Analysis (VTA), a novel framework that combine verbal and latent reasoning to produce stock time-series forecasts that are both accurate and interpretable. To reason over time-series, we convert stock price data into textual annotations and optimize the reasoning trace using an inverse Mean Squared Error (MSE) reward objective. To produce time-series outputs from textual reasoning, we condition the outputs of a time-series backbone model on the reasoning-based attributes.…
Peer Reviews
Decision·ICLR 2026 Poster
The paper proposes a novel framework, Verbal Technical Analysis (VTA), which effectively integrates large language models with time-series forecasting to produce both accurate and interpretable stock predictions. It introduces an innovative reinforcement learning objective (Time-GRPO) using inverse MSE as a reward to improve the quality of verbal reasoning aligned with forecast accuracy. The model generates natural language reasoning traces that are rated highly by financial domain experts, ad
The proposed training pipeline is multi-staged and complex, involving cold-start RL, supervised fine-tuning, and joint conditional training, which may limit reproducibility and increase implementation overhead. The paper does not compare VTA’s reasoning or performance against human-crafted technical analysis rules or human analysts, which would be a natural baseline for such a task. The paper does not compare VTA with financial time seires LLMs, e.g., Kronous, which are related to financial an
- Well-written, easy to understand - This paper contrasts with others in the area by combining neural representations of LMs with text embeddings. Most other papers focus only on time series expressed as text or on neural embeddings, not both. - Strong choice of baselines and third-party datasets - The use of experts for annotation is useful and lends confidence towards the model's performance
On the whole I think this is a very strong paper. However, without details on how baselines were tuned it's difficult for me to have confidence in the method's performance. If the authors can provide this information then I'll happily increase my score to an accept. To be clear, my primary concern is that the proposed method could be outperformed by (1) simple statistical methods or (2) strongly hyperparameter-tuned baselines.
pros: 1. Interesting conceptual direction — bridging LLM reasoning and numerical forecasting is a timely and under-explored research frontier, particularly for financial data. 2. Clear problem motivation — the authors highlight the lack of interpretable time-series reasoning methods and position VTA as a step toward explainable financial LLMs. 3. Engineering completeness — the framework, ablations, and appendices (notably the extensive tables on pages 6–9 and expert evaluation in Appendix C) a
1. The “Time-GRPO” formulation largely reuses GRPO (Shao et al., 2024) with an inverse-MSE reward; conceptually it is a direct extension rather than a fundamentally new RL objective. The “joint conditional training” (Figure 3) borrows heavily from classifier-free guidance (Ho & Salimans 2022) and conditional diffusion principles. The novelty lies more in application assembly than in algorithmic innovation. 2. The paper never establishes whether the verbal reasoning causally improves forecasting
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsStock Market Forecasting Methods · Forecasting Techniques and Applications · Time Series Analysis and Forecasting