Skip to content

Reasoning on Time-Series for Financial Technical Analysis

Conference: ICLR2026 arXiv: 2511.08616 Code: chen-jan/VTA Area: Time Series Keywords: Time series reasoning, financial technical analysis, reinforcement learning, LLM fine-tuning, interpretable forecasting Authors: Kelvin J.L. Koa, Jan Chen, Yunshan Ma, Huanhuan Zheng, Tat-Seng Chua (NUS, TUM, SMU, CityU HK)

TL;DR

This paper proposes the Verbal Technical Analysis (VTA) framework, which combines the linguistic reasoning capabilities of LLMs with the pattern-capturing capacity of time-series models. Time-GRPO reinforcement learning is employed to optimize reasoning chains, and inferred attributes are used to condition time-series forecasting, achieving financial time-series prediction that is both accurate and interpretable.

Background & Motivation

Limitations of LLMs in finance: Existing financial LLMs primarily analyze textual reports (earnings Q&A, sentiment analysis) while neglecting interpretable analysis of historical price data — i.e., Technical Analysis (TA) — which is critically important for trading practitioners.

LLMs are poor at time-series reasoning: Prior work (Merrill et al., 2024) has shown that LLMs perform "remarkably bad" at zero-shot time-series reasoning, and feeding raw time-series data directly yields poor results.

Time-series LLMs sacrifice interpretability: Methods such as Time-LLM and CALF modify the embedding space to produce time-series forecasts, but in doing so LLMs lose their natural language reasoning capability and cannot provide interpretable analysis.

Inadequacy of existing interpretable approaches: The most closely related work, TimeCAP, produces only classification label predictions rather than full time-series trajectories, and its reasoning relies on external auxiliary data rather than endogenous signals.

Cross-domain challenge: The task requires switching between two domains — the time-series domain (stock prices) for inputs and outputs, and the natural language domain for the reasoning process — which increases modeling difficulty.

Inherently interpretable signals in financial time series: Unlike general time series, financial data contains a wealth of expert-studied technical indicators (MACD, RSI, Bollinger Bands, etc.), which provide a natural anchor for language-based reasoning.

Method

Overall Architecture

The VTA framework consists of three core components:

  • Time-Series Reasoning: Teaches the LLM to perform linguistic reasoning over time-series inputs.
  • Time-Series Forecasting: Uses a backbone time-series model to capture underlying complex patterns.
  • Joint Conditional Training: Injects reasoning attributes as conditions into the time-series forecasting process.

Problem Formulation

Given a historical input of \(T\) trading days \(\mathbf{X} = \{\mathbf{x}_{t-T+1}, \ldots, \mathbf{x}_t\}\), where \(\mathbf{x}_t = [o_t, h_t, l_t, v_t, c_t, p_t]\) (open, high, low, volume, close, and adjusted close price), the goal is to generate:

  • A linguistic reasoning trace \(\mathbf{v}\)
  • Price forecasts \(\mathbf{y} = \{p_{t+1}, \ldots, p_{t+T'}\}\) for the next \(T'\) trading days

Experiments use \(T = T' = 10\) (short-term forecasting).

Time-Series Reasoning (Time-GRPO)

Textual annotation: Raw time-series data is converted into a textual annotation \(\mathbf{X'} = \mathbf{f}(\mathbf{X})\), including statistical summaries (mean/min/max) and financial technical indicators (moving averages, momentum, MACD, RSI, Bollinger Bands, etc.).

Time-GRPO objective: Adapted from GRPO (Shao et al., 2024), with the core formulation:

\[\mathcal{L}_{\text{time-grpo}}(\theta) = \mathbb{E}_{\mathbf{q} \sim \mathcal{Q}} \frac{1}{G} \sum_{i=1}^{G} \left( \min\left(\frac{\pi_\theta(\mathbf{o_i}|\mathbf{q})}{\pi_{\theta_{\text{old}}}(\mathbf{o_i}|\mathbf{q})} A_i, \text{clip}(\cdot, 1{-}\epsilon, 1{+}\epsilon) A_i \right) - \beta \mathbb{D}_{\text{KL}}(\pi_\theta \| \pi_{\text{ref}}) \right)\]

Inverse MSE reward: Encourages reasoning chains to maximize prediction accuracy:

\[r_{\text{MSE}}(\theta) = \frac{1}{\lambda \cdot \|\hat{\mathbf{y}}_\theta - \mathbf{y}\|_2^2}\]

The inverse MSE formulation is used because the reward must be maximized (smaller MSE yields a larger reward).

Multi-stage training pipeline:

  1. Cold-Start stage: Time-GRPO is applied to generate initial training samples; performance gains are limited at this stage but the generated data serves as the foundation for subsequent stages.
  2. Rejection Sampling + SFT stage: Reasoning chains in the top 10% MSE within each bucket are retained for supervised fine-tuning.
  3. RL Optimization stage: Time-GRPO is applied again on the model that has already learned to reason, to search for the optimal reasoning strategy.

Time-Series Forecasting Backbone

Based on GPT-2, adapted via cross-modal fine-tuning:

  • Time-series inputs are processed through Embedding + Multi-head Attention to produce projected time tokens \(\mathbf{X}_{\text{time}}\).
  • PCA is applied to LLM word embeddings to obtain principal component word embeddings \(\hat{\mathbf{D}}\).
  • Multi-head Cross-Attention aligns time tokens with word embeddings:
\[\mathbf{X}_{\text{text}} = \text{Softmax}\left(\frac{\mathbf{Q}\mathbf{K}^\top}{\sqrt{C}}\right)\mathbf{V}\]
  • Layer-wise feature regularization aligns the temporal and text branches:
\[\mathcal{L}_{\text{feature}} = \sum_{n=1}^{N} \gamma^{(N-n)} \text{sim}\left(\phi_{\text{text}}^n(\mathbf{F}_{\text{text}}^n), \phi_{\text{time}}^n(\mathbf{F}_{\text{time}}^n)\right)\]

Joint Conditional Training

Descriptive attribute classes \(\mathbf{c}\) (maximum/minimum/mean values) are extracted from the reasoning output and used to condition the time-series forecasting:

\[\mathcal{L}_{\text{forecast}}(\phi) = \mathbb{E}_{\mathbf{X}, \mathbf{y}, \mathbf{c}} \left[\|\hat{\mathbf{y}}_\psi(\mathbf{X}, \tilde{\mathbf{c}}) - \mathbf{y}\|^2\right]\]

With probability \(p_{\text{uncond}}=0.3\), \(\mathbf{c}\) is randomly set to null (analogous to Classifier-Free Guidance), jointly training both conditional and unconditional paths. At inference:

\[\hat{\mathbf{y}} = s \cdot \hat{\mathbf{y}}_\psi(\mathbf{X}, \mathbf{c}) + (1-s) \cdot \hat{\mathbf{y}}_\theta(\mathbf{X})\]

where the guidance scale \(s=0.1\).

Key Experimental Results

Main Results: Forecasting Performance Comparison

Datasets: ACL18 StockNet (88 US stocks, 2012–2017) + Dow Jones / China A50 / EURO STOXX 50 (2024)

Model StockNet MSE StockNet MAE All MSE All MAE
GPT-4.1 mini 0.0846 0.1827 0.2014 0.2376
DeepSeek-R1 0.0788 0.1853 0.1428 0.2323
TimesNet 0.0708 0.1789 0.1286 0.2229
TimeLLM 0.0704 0.1780 0.1262 0.2210
CALF 0.0674 0.1738 0.1235 0.2180
VTA (Ours) 0.0659 0.1701 0.1178 0.2122

VTA achieves the best MSE and MAE across all four datasets, with an overall MSE improvement of 4.6% and MAE improvement of 2.7%.

Ablation Study: Contribution of the Multi-Stage Pipeline

Stage Llama-3.1-8B MSE Qwen-2.5-3B MSE Qwen-2.5-7B MSE
Base Model 0.1482 0.1707 0.0949
Cold Start (RL) 0.1475 0.1648 0.0941
SFT for Reasoning 0.1168 0.1032 0.0893
RL for Reasoning 0.0955 0.0832 0.0686
+ Conditioning (VTA) 0.0667 0.0672 0.0659

Key Findings: - Cold-Start RL yields only a 1.6% average improvement but generates the data essential for subsequent stages. - Rejection sampling + SFT followed by RL achieves a 20.3% improvement, validating the effectiveness of the multi-stage pipeline. - Conditioning the backbone model further reduces error, demonstrating the complementary benefit of combining external reasoning with internal pattern modeling. - Qwen-2.5-7B performs best as the reasoning model, but with conditioning training the 3B model achieves comparable performance.

Reasoning Quality Evaluation

25 financial industry experts from JPMorgan, UBS, Evercore, Allianz, and other institutions conducted blind evaluations (1–5 scale) of reasoning chains from VTA, GPT-4.1 mini, and DeepSeek-R1:

  • Depth, Accuracy, Relevance: VTA leads significantly, reflecting its effective use of technical indicators and reasoning capacity.
  • Coherence, Clarity: Gaps are smaller, as general-purpose LLMs retain inherent advantages in text fluency.

Portfolio Evaluation

Model Returns Volatility Max Drawdown Sharpe Ratio
TimeLLM 0.2185 0.1193 -0.1040 1.5230
CALF 0.2019 0.1247 -0.0981 1.4566
VTA (Ours) 0.2409 0.1185 -0.0883 1.7190

VTA substantially outperforms baselines on Sharpe Ratio (1.7190 vs. the second-best 1.5230), confirming its practical value in real investment scenarios.

Reasoning Perturbation Experiments

  • Removing technical indicators leads to a noticeable drop in forecasting performance, indicating that the reasoning chain provides genuinely useful guidance signals.
  • Adding adversarial noise degrades performance but with inconsistent trends, possibly because the model learns during joint training to rely more heavily on the time-series backbone when reasoning is unreliable.

Highlights & Insights

  1. Elegant cross-domain bridging: Financial technical indicators serve as a natural bridge between the time-series and language domains, addressing the known weakness of LLMs in processing raw time-series data directly.
  2. Time-GRPO design: Using inverse MSE as the RL reward directly drives reasoning chain optimization toward prediction accuracy, requiring no manual annotation of reasoning data.
  3. Multi-stage training pipeline: The progressive design of Cold-Start → Rejection Sampling SFT → RL yields more stable and efficient training.
  4. Transfer of Classifier-Free Guidance: The conditional guidance technique from diffusion models is adapted for time-series forecasting, jointly training conditional and unconditional paths.
  5. Comprehensive evaluation: Evaluation covers not only prediction accuracy but also expert reasoning quality scores and Markowitz portfolio validation.

Limitations & Future Work

  1. Restricted to financial time series: Cross-domain experiments (medical/energy) indicate that VTA's reasoning advantage relies on the inherently interpretable signals of financial technical indicators; on general time-series data, the approach degrades to simple trend extrapolation.
  2. Short-horizon forecasting: \(T=T'=10\) covers only short-term trading scenarios; the effectiveness of long-horizon forecasting remains unvalidated.
  3. Alignment between reasoning and forecasting: Conditioning uses only simple attributes (max/min/mean); richer information within the reasoning chain (trend direction, indicator signals) is not fully exploited.
  4. Fixed guidance scale: \(s=0.1\) indicates that the model relies primarily on the backbone in practice, and the actual contribution of reasoning guidance is relatively small.
  5. Computational cost: The multi-stage RL + LLM reasoning + time-series backbone joint training pipeline entails considerable resource consumption.
  6. Base model selection: Only three LLM bases are tested (Llama-3.1-8B, Qwen-2.5-3B/7B); larger-scale models remain unexplored.
Direction Representative Work Distinction from VTA
Financial LLMs Fin-R1, FinMem, SEP Analyze textual reports/news; do not process price time series
Time-series LLMs Time-LLM, CALF Modify embedding space; lose language reasoning capability
Time-series reasoning TimeCAP Relies on external auxiliary data; produces only classification labels
LLM time-series reasoning Merrill et al. Identifies poor zero-shot TS reasoning in LLMs; VTA addresses this via RL fine-tuning
Reasoning optimization DeepSeek-R1, GRPO VTA adapts GRPO into Time-GRPO with an inverse MSE reward

Rating

  • Novelty: ⭐⭐⭐⭐ — Combining RL-based reasoning optimization (GRPO) with time-series forecasting, and transferring Classifier-Free Guidance to time-series conditioning, are genuinely novel contributions.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ — Four datasets, 14+ baselines, ablation studies, expert evaluations, portfolio validation, and cross-domain generalization analysis constitute an exceptionally comprehensive evaluation.
  • Writing Quality: ⭐⭐⭐⭐ — Well-structured, with well-motivated problem formulation and rich figures and tables.
  • Value: ⭐⭐⭐⭐ — Interpretable financial forecasting offers direct practical value to practitioners; Sharpe Ratio validation confirms real investment potential.