Paper: arXiv 2306.11025

Authors: Xinli Yu, Zheng Chen, Yuan Ling, Shujing Dong, Zongyi Liu, Yanbin Lu

Abstract

This paper presents a novel study on harnessing Large Language Models’ (LLMs) outstanding knowledge and reasoning abilities for explainable financial time series forecasting. The application of machine learning models to financial time series comes with several challenges, including the difficulty in cross-sequence reasoning and inference, the hurdle of incorporating multi-modal signals from historical news, financial knowledge graphs, etc., and the issue of interpreting and explaining the model results. In this paper, we focus on NASDAQ-100 stocks, making use of publicly accessible historical stock price data, company metadata, and historical economic/financial news. We conduct experiments to illustrate the potential of LLMs in offering a unified solution to the aforementioned challenges. Our experiments include trying zero-shot/few-shot inference with GPT-4 and instruction-based fine-tuning with a public LLM model Open LLaMA. We demonstrate our approach outperforms a few baselines, including the widely applied classic ARMA-GARCH model and a gradient-boosting tree model. Through the performance comparison results and a few examples, we find LLMs can make a well-thought decision by reasoning over information from both textual news and price time series and extracting insights, leveraging cross-sequence information, and utilizing the inherent knowledge embedded within the LLM. Additionally, we show that a publicly available LLM such as Open-LLaMA, after fine-tuning, can comprehend the instruction to generate explainable forecasts and achieve reasonable performance, albeit relatively inferior in comparison to GPT-4.

Complexity vs Empirical Score

  • Math Complexity: 4.0/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper relies on high-level machine learning concepts and LLM architectures rather than dense mathematical derivations, but features a rigorous empirical setup with real-world NASDAQ-100 data, multiple baselines (ARMA-GARCH, gradient boosting), and detailed performance comparisons.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Use LLMs for Explainable<br>Financial Time Series Forecasting"] --> B["Methodology<br>LLM Inference & Fine-tuning<br>(GPT-4 & Open-LLaMA)"]
  
  B --> C["Input Data<br>Historical Prices +<br>News & Knowledge Graphs"]
  
  C --> D["Processing<br>Unified Analysis<br>(Multi-modal Reasoning)"]
  
  D --> E["Comparative Analysis<br>vs. Baselines<br>(ARMA-GARCH, GBM)"]
  
  E --> F["Key Findings<br>• LLMs outperform baselines<br>• Unified solution for multi-modal data<br>• Explainable reasoning process"]

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 using large language models with structured prompts and fine‑tuning to generate explainable forecasts for NASDAQ‑100 stock returns.

Source sentence
how LLMs, specifically GPT-4 and Open LLaMA, can be utilized to generate forecasts and provide human‑readable explanations for their predictions.

Assumes 4

  • Price series are transformed into percentage‑change bins for forecasting. data
    Source sentence
    We first normalize the numeric price time series as a percentage‑change time series, and then categorize the percentage change into bins.
  • Weekly and monthly returns are defined as the percentage change from the beginning to the end of the week/month. data
    Source sentence
    forecasting weekly/monthly stock returns (defined as the percentage change in stock price from the beginning to the end of the week/month)
  • Zero‑shot/few‑shot inference with GPT‑4 can be performed without model fine‑tuning. model
    Source sentence
    We demonstrate our structured design of prompts for LLMs and apply the state‑of‑the‑art GPT‑4 model
  • Open LLaMA must be fine‑tuned because it cannot handle instruction‑based prompts in zero‑shot mode. model
    Source sentence
    zero‑shot inference, typically tends to replicate portions of the prompt rather than executing the instructions

Data

  • NASDAQ‑100 daily stock price data · Yahoo Finance market-prices public
    NASDAQ‑100 constituents
    Source sentence
    price data from Yahoo Finance using yfinance package.
  • Company metadata fundamentals public
    NASDAQ‑100 constituents
    Source sentence
    supplemented by metadata about the stock company
  • Historical economic/financial news text-news public
    NASDAQ‑100 constituents
    Source sentence
    concerning both the specific stock and the broader financial/economic landscape

Recognised sources: Alpha Vantage Yahoo Finance

Costs

Not applicable The paper reports no trading or execution result.

Finds 3

  • LLM‑based forecasting outperforms classic ARMA‑GARCH and gradient‑boosting tree baselines. performance
    Source sentence
    We demonstrate our approach outperforms a few baselines, including the widely applied classic ARMA‑GARCH model and a gradient‑boosting tree model.
  • Fine‑tuned Open LLaMA achieves reasonable performance but is inferior to GPT‑4. performance
    Source sentence
    a publicly available LLM such as Open‑LLaMA, after fine‑tuning, can comprehend the instruction to generate explainable forecasts and achieve reasonable performance, albeit relatively inferior in comparison to GPT‑4.
  • Incorporating Chain‑of‑Thought reasoning significantly improves LLM forecasting performance. performance
    Source sentence
    Notably, the integration of a step‑by‑step reasoning process, inspired by the Chain of Thought (COT) approach, significantly enhanced the performance of our LLM‑based models.

Checked by others

Qualified by

  • Large Language Models and Stock Investing: Is the Human Factor Required? (2026)
    On the question: Does chain‑of‑thought reasoning improve LLM financial forecasting performance?
    Compare the findings

    This paper: Incorporating Chain‑of‑Thought reasoning significantly improves LLM forecasting performance.

    That paper: Despite improvements, reasoning failures remain pervasive, leading to unstable risk‑adjusted performance and occasional significant mistakes.

    A reports significant improvement, while B finds pervasive reasoning failures leading to unstable risk‑adjusted results.

Cited by 22 papers in the archive (20 by independent authors): Time Series Foundation Models for Multivariate Financial … (discusses its findings); Learning to Generate Explainable Stock Predictions using … (uses its method or data)

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.