Paper: arXiv 2510.01203

Authors: Lokesh Antony Kadiyala, Amir Mirzaeinia

Abstract

The stock market is extremely difficult to predict in the short term due to high market volatility, changes caused by news, and the non-linear nature of the financial time series. This research proposes a novel framework for improving minute-level prediction accuracy using semantic sentiment scores from top ten different large language models (LLMs) combined with minute interval intraday stock price data. We systematically constructed a time-aligned dataset of AAPL news articles and 1-minute Apple Inc. (AAPL) stock prices for the dates of April 4 to May 2, 2025. The sentiment analysis was achieved using the DeepSeek-V3, GPT variants, LLaMA, Claude, Gemini, Qwen, and Mistral models through their APIs. Each article obtained sentiment scores from all ten LLMs, which were scaled to a [“0, 1”] range and combined with prices and technical indicators like RSI, ROC, and Bollinger Band Width. Two state-of-the-art such as Reformer and Mamba were trained separately on the dataset using the sentiment scores produced by each LLM as input. Hyper parameters were optimized by means of Optuna and were evaluated through a 3-day evaluation period. Reformer had mean squared error (MSE) or the evaluation metrics, and it should be noted that Mamba performed not only faster but also better than Reformer for every LLM across the 10 LLMs tested. Mamba performed best with LLaMA 3.3–70B, with the lowest error of 0.137. While Reformer could capture broader trends within the data, the model appeared to over smooth sudden changes by the LLMs. This study highlights the potential of integrating LLM-based semantic analysis paired with efficient temporal modeling to enhance real-time financial forecasting.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced deep learning architectures (state-space models, transformers with locality-sensitive hashing) and discusses the theoretical efficiency of Mamba (linear complexity) and Reformer (LSH), indicating high mathematical density. It also features a highly empirical setup with a custom, time-aligned minute-level dataset, hyperparameter tuning via Optuna, and clear evaluation metrics (MSE) on a hold-out test set, making it backtest-ready.

Research Flowchart

  flowchart TD
  A["Research Goal: Improve minute-level stock prediction<br>using LLM sentiments & price data"] --> B["Data Collection & Processing"]
  B --> C["Sentiment Analysis<br>via Top 10 LLMs"]
  B --> D["Price & Technical Indicators<br>RSI, ROC, Bollinger"]
  C & D --> E["Model Training<br>Optuna Hyperparameter Optimization"]
  E --> F{"Evaluation on 3-Day Period"}
  F --> G["Mamba Architecture<br>MSE: 0.137 (Best with LLaMA 3.3-70B)<br>Faster & More Accurate"]
  F --> H["Reformer Architecture<br>Over-smoothed sudden market changes"]

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 a novel framework that integrates sentiment scores from ten large language models with minute‑level intraday stock price data to improve short‑term stock prediction, and evaluates Mamba and Reformer models on this data.

Source sentence
This research proposes a novel framework for improving minute-level prediction accuracy using semantic sentiment scores from top ten different large language models (LLMs) combined with minute interval intraday stock price data.

Assumes 5

  • Only Apple Inc. (AAPL) stock is considered for prediction. market
    Source sentence
    We systematically constructed a time-aligned dataset of AAPL news articles and 1-minute Apple Inc. (AAPL) stock prices for the dates of April 4 to May 2, 2025.
  • News articles published outside regular market hours are shifted to the next trading day at 9:30 AM EST. data
    Source sentence
    Articles that were published outside regular business hours (before 9:30 AM or after 4:00 PM ET), as well as those from non-business days
  • Sentiment scores from each LLM are scaled to a [0, 1] range before being used as model inputs. data
    Source sentence
    Each article obtained sentiment scores from all ten LLMs, which were scaled to a [0, 1] range and combined with prices and technical indicators...
  • Technical indicators (RSI, ROC, Bollinger Band Width) are computed from the 1‑minute price series and are treated as stationary inputs. model
    Source sentence
    combined with prices and technical indicators like RSI, ROC, and Bollinger Band Width.
Show 1 more
  • Model hyper‑parameters are tuned using Optuna over 20 iterations and evaluated on a 3‑day hold‑out period. model
    Source sentence
    Hyper parameters were optimized by means of Optuna and were evaluated through a 3-day evaluation period.

Data

Backtest sample: 2025-04 – 2025-05 · Apple Inc. (AAPL) intraday stock prices · equities · intraday

  • AAPL news articles · News API text-news public
    2025-04-04 to 2025-05-02 · Apple Inc.
    Source sentence
    The News API was used to fetch news headlines about Apple Inc. (AAPL) that were published between April 4, 2025, at 4:01 PM Eastern Time (ET), and May 2, 2025, at 4:00 PM ET.
  • AAPL 1‑minute price data · Polygon.io API market-prices commercial
    2025-04-04 to 2025-05-02 · Apple Inc.
    Source sentence
    High-frequency stock price data at 1‑minute intervals during U.S. market hours was collected using the Polygon.io API.

Recognised sources: Polygon.io

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • Mamba achieved the lowest error of 0.137 when using LLaMA 3.3‑70B sentiment scores, outperforming Reformer across all ten LLMs. performance
    Source sentence
    Mamba performed best with LLaMA 3.3–70B, with the lowest error of 0.137.
  • Mamba repeatedly outperformed Reformer on 9 out of 10 LLM‑based inputs, showing lower MSE and better short‑term adaptability. performance
    Source sentence
    Mamba repeatedly outperformed Reformer across 9 out of 10 LLM-based inputs.
  • Reformer’s best per‑performance was with Qwen Turbo, yielding an MSE of 2.647, but it was less accurate than the top Mamba models. performance
    Source sentence
    The Qwen Turbo setup had the best per‑formance in the Reformer model, with an MSE value of 2.647.
  • Mamba demonstrated greater responsiveness to rapid price shifts, maintaining tight alignment with ground truth, whereas Reformer produced smoother curves. predictive-accuracy
    Source sentence
    The Mamba model displayed greater responsiveness to minor market trends and price surges, maintaining tight alignment with ground truth data during rapid price shifts.
  • Despite higher error margins, Reformer captured broader trend directions reasonably well when paired with certain LLMs (e.g., Qwen Turbo, DeepSeek V3). predictive-accuracy
    Source sentence
    showed reasonable outcomes when paired with Qwen Turbo and DeepSeek V3, capturing the broader structure of stock market trends.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 46 related papers). Coverage is partial, so this is neither replication nor refutation.

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.

Cited References (12)

  1. Mamba: Linear-Time Sequence Modeling with Selective State Spaces (2023) — link
  2. Reformer: The Efficient Transformer (2020) — link
  3. Optuna: A Next-generation Hyperparameter Optimization Framework (2019) — link
  4. Decoupled Weight Decay Regularization (2017)
  5. Scikit-learn: Machine Learning in Python (2011) — link
  6. “DashScope API for Qwen Turbo,” (2025)
  7. “REST API for real-time news articles,”
  8. “Financial Market Data APIs,” 2025
and 4 more
  1. Percentage Change in Closing Price ( % ∆ Close): represents trends in short-term returns based on minute-level variations
  2. API, “Unified API access to LLMs including GPT, Claude, LLaMA, Mistral,” 2025
  3. “DeepSeek-V3 Language Model API,”
  4. Price Difference ( ∆ Close): Represents the absolute difference between consecutive close prices

0 of 12 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.