Paper: arXiv 2511.08616
Authors: Kelvin J. L. Koa, Jan Chen, Yunshan Ma, Huanhuan Zheng, Tat-Seng Chua
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. Experiments on stock datasets across U.S., Chinese, and European markets show that VTA achieves state-of-the-art forecasting accuracy, while the reasoning traces also perform well on evaluation by industry experts.
Complexity vs Empirical Score
- Math Complexity: 6.5/10
- Empirical Rigor: 6.0/10 (capped from 8.0: transaction costs not modelled; see penalty flags)
- Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 8.0)
Why this score: The paper introduces a novel reinforcement learning objective (Time-GRPO) with clipped ratios and KL-divergence regularization, and conditions forecasts on reasoning attributes, showing moderate-to-high mathematical density. It also validates the model across U.S., Chinese, and European stock datasets, performing investment simulations and expert evaluations, demonstrating significant empirical validation and implementation focus.
Research Flowchart
flowchart TD
A["Research Goal: Combine verbal & latent reasoning<br>for accurate & interpretable stock forecasts"] --> B
subgraph B ["Methodology: Verbal Technical Analysis VTA"]
B1["Input: Stock Price Time-Series Data"] --> B2["Convert to Textual Annotations"]
B2 --> B3["Verbal Reasoning via LLMs"]
B3 --> B4["Inverse MSE Reward Optimization"]
B4 --> B5["Conditional Time-Series Generation"]
end
B --> C["Output: Stock Price Forecasts & Reasoning Traces"]
C --> D["Outcome: State-of-the-Art Accuracy<br>Interpretability validated by Experts<br>Across US, China, Europe Markets"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes the Verbal Technical Analysis (VTA) framework that combines a backbone time-series model with a reasoning LLM to produce interpretable stock forecasts.
Source sentence
we propose our Verbal Technical Analysis (VTA) framework, which combines a backbone time-series model (which we termed as “latent thinking”) with a reasoning LLM (termed as “verbal reasoning”) to produce interpretable stock time-series forecasts.
Assumes 4
- The conditional probability puncond is set to 0.3 and the guidance scale s to 0.1 for the model. model
Source sentence
For the forecasting model, we use GPT-2 (Radford et al., 2019) as the base model. For hyperparameters, we set the conditional probability puncond to 0.3 and the guidance scale s to 0.1.
- Stocks are selected bias‑free from well‑known indices (Dow Jones, FTSE China A50, EURO STOXX 50). data
Source sentence
To ensure a bias-free selection, we choose the stocks from well-known indices, i.e., the Dow Jones, the FTSE China A50 Index and the EURO STOXX 50.
- The forecasting loss is mean‑squared error between predicted and ground‑truth series. model
Source sentence
The model is then fine-tuned using MSE loss with the ground truth series y.
- Portfolio construction uses Markowitz optimization over the 10‑day predictions. execution
Source sentence
This is achieved by performing Markowitz optimization (Markowitz, 1952) across the 10-day predictions.
Data
global equity markets (U.S., Chinese, European) · equities · daily
- ACL18 StockNet dataset benchmark public
Source sentence
The first is the ACL18 StockNet dataset (Xu & Cohen, 2018), which includes historical price data for 88 U.S. stocks that are selected to represent the top 8-10 companies by market capitalization in each of the major industries. The dataset spans the period of 01/09/2012 to 01/09/2017.
- U.S., Chinese, and European market data market-prices unknown
Source sentence
To further show the generalization ability of the model, we also collect additional stock data from across the U.S., Chinese, and European markets for testing.
Costs
Not modeled
Finds 3
Sharpe 1.72 · max drawdown 8.8%
- VTA achieves the best performance among all models in both MSE and MAE on the test sets. performance
Source sentence
Our proposed VTA model demonstrates the best performance in stock forecasting in both MSE and MAE.
- VTA obtains the highest Sharpe ratio among all evaluated models in portfolio optimization. performance
Source sentence
VTA achieves the highest Sharpe ratio among all models, which represents the risk-adjusted returns.
- The second RL fine‑tuning stage improves average MSE by about 20.3% over the base model. statistical
Source sentence
the second RL fine-tuning, which uses the same Time-GRPO objective, produces an average improvement of 20.3%.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 31 related papers). Coverage is partial, so this is neither replication nor refutation.
1 paper in the archive
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (95)
- Newton’s financial misadventures in the South Sea Bubble (2019)
- Fortune’s formula: The untold story of the scientific betting system that beat the casinos and Wall Street (2010)
- Predict Stock Prices Using RNN: Part 1 (2017) — link
- Temporal relational ranking for stock prediction (2019)
- Diffusion variational autoencoder for tackling stochasticity in multi-step regression stock price prediction — → 2309.00073
- Human-Centric Justification of Machine Learning Predictions. (2017)
- Technical analysis: the complete resource for financial market technicians (2010)
- Fin-R1: A Large Language Model for Financial Reasoning through Reinforcement Learning (2025) — link
and 87 more
- Pixiu: A large language model, instruction data and evaluation benchmark for finance (2023) — link
- Time-LLM : Time series forecasting by reprogramming large language models (2024)
- TimeCAP: Learning to Contextualize, Augment, and Predict Time Series Events with Large Language Model Agents (2025)
- Technical analysis of the financial markets: A comprehensive guide to trading methods and applications (1999)
- The combined signal approach to technical analysis: A review & commentary (2009)
- The profitability of technical analysis: A review (2004)
- Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models (2023)
- Making large language models perform better in knowledge graph completion (2024)
- Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning (2025) — link
- Deepseekmath: Pushing the limits of mathematical reasoning in open language models (2024) — link
- Decoding Time Series with LLMs: A Multi-Agent Framework for Cross-Domain Annotation (2024) — link
- Learning multiple stock trading patterns with temporal routing adaptor and optimal transport (2021)
- MATCC: A Novel Approach for Robust Stock Price Prediction Incorporating Market Trends and Cross-time Correlations (2024)
- Period-aggregated transformer for learning latent seasonalities in long-horizon financial time series (2024)
- An enhanced Transformer framework with incremental learning for online stock price prediction (2025)
- Multi-Sensor Temporal Fusion Transformer for Stock Performance Prediction: An Adaptive Sharpe Ratio Approach (2025)
- Listening to chaotic whispers: A deep learning framework for news-oriented stock trend prediction (2018)
- Numhtml: Numeric-oriented hierarchical transformer model for multi-task financial forecasting (2022)
- Bloomberggpt: A large language model for finance (2023) — link
- Learning to generate explainable stock predictions using self-reflective large language models — → 2402.03659
- TAT-QA: A question answering benchmark on a hybrid of tabular and textual content in finance (2021)
- Promptcast: A new prompt-based learning paradigm for time series forecasting (2023)
- From news to forecast: Integrating event analysis in llm-based time series forecasting with reflection (2024)
- Chronos: Learning the language of time series (2024)
- Large language models are zero-shot time series forecasters (2023)
- S\textasciicircum 2 IP-LLM: Semantic Space Informed Prompt Learning with LLM for Time Series Forecasting (2024)
- Classifier-free diffusion guidance (2022) — link
- A time series is worth 64 words: Long-term forecasting with transformers (2022) — link
- Diffusion models beat gans on image synthesis (2021)
- Stock movement prediction from tweets and historical prices (2018)
- Fino1: On the transferability of reasoning enhanced llms to finance (2025) — link
- Deepseek-vl: towards real-world vision-language understanding (2024) — link
- In-context instruction learning (2023) — link
- Calf: Aligning llms for time series forecasting via cross-modal fine-tuning (2025)
- One fits all: Power general time series analysis by pretrained lm (2023)
- Llm4ts: Two-stage fine-tuning for time-series forecasting with pre-trained llms (2023)
- Tempo: Prompt-based generative pre-trained transformer for time series forecasting (2023) — link
- Test: Text prototype aligned embedding to activate llm’s ability for time series (2023) — link
- A simple framework for contrastive learning of visual representations (2020)
- PEN: prediction-explanation network to forecast stock price movement with better explainability (2023)
- DHMoE: Diffusion Generated Hierarchical Multi-Granular Expertise for Stock Prediction (2025)
- Introducing GPT-4.1 in the API (2025) — link
- Informer: Beyond efficient transformer for long sequence time-series forecasting (2021)
- Attention is all you need (2017)
- Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting (2023)
- Reformer: The efficient transformer (2020) — link
- LightTS: Lightweight time series classification with adaptive ensemble distillation (2023)
- Are transformers effective for time series forecasting? (2023)
- Film: Frequency improved legendre memory model for long-term time series forecasting (2022)
- Non-stationary transformers: Exploring the stationarity in time series forecasting (2022)
- Micn: Multi-scale local and global context modeling for long-term series forecasting (2023)
- Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting (2021)
- Timesnet: Temporal 2d-variation modeling for general time series analysis (2022) — link
- Tsmixer: An all-mlp architecture for time series forecasting (2023) — link
- Lora: Low-rank adaptation of large language models. (2022)
- Language models are unsupervised multitask learners (2019)
- Adam: A method for stochastic optimization (2014) — link
- Enhancing stock movement prediction with adversarial training (2018) — link
- Time horizon-aware modeling of financial texts for stock price prediction (2021)
- Deep attentive learning for stock movement prediction from social media text and company correlations (2020)
- Qwen2.5: A Party of Foundation Models (2024) — link
- Technical analysis (1995)
- Generative time series forecasting with diffusion, denoise, and disentanglement (2022)
- Judging llm-as-a-judge with mt-bench and chatbot arena (2023)
- A survey on llm-as-a-judge (2024) — link
- Portfolio Selection (1952) — link
- The sharpe ratio (1994)
- Principles (2018)
- A random walk down Wall Street: the time-tested strategy for successful investing (2019)
- Adding conditional control to text-to-image diffusion models (2023)
- A dual-stage attention-based recurrent neural network for time series prediction (2017) — link
- Predicting stock prices using LSTM (2017)
- Towards time series reasoning with llms (2024) — link
- Position: What can large language models tell us about time series analysis (2024)
- Language models still struggle to zero-shot reason about time series (2024) — link
- Can LLMs Understand Time Series Anomalies? (2024) — link
- Generative adversarial text to image synthesis (2016)
- High-resolution image synthesis with latent diffusion models (2022)
- Justice or prejudice? quantifying biases in llm-as-a-judge (2024) — link
- The Black Swan: The Impact of the Highly Improbable (2007)
- Finmem: A performance-enhanced llm trading agent with layered memory and character design — → 2311.13743
- Fincon: A synthesized llm multi-agent system with conceptual verbal reinforcement for enhanced financial decision making (2024)
- Temporal data meets LLM–explainable financial time series forecasting — → 2306.11025
- Training language models to follow instructions with human feedback (2022)
- Time-mqa: Time series multi-task question answering with context enhancement (2025) — link
- Detecting influenza epidemics using search engine query data (2009)
- Are language models actually useful for time series forecasting? (2024)
4 of 95 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.