Paper: arXiv 2502.07393

Authors: Mostapha Benhenda

Abstract

This paper presents a novel risk-sensitive trading agent combining reinforcement learning and large language models (LLMs). We extend the Conditional Value-at-Risk Proximal Policy Optimization (CPPO) algorithm, by adding risk assessment and trading recommendation signals generated by a LLM from financial news. Our approach is backtested on the Nasdaq-100 index benchmark, using financial news data from the FNSPID dataset and the DeepSeek V3, Qwen 2.5 and Llama 3.3 language models. The code, data, and trading agents are available at: https://github.com/benstaf/FinRL_DeepSeek

Complexity vs Empirical Score

  • Math Complexity: 6.5/10
  • Empirical Rigor: 6.0/10 (capped from 7.0: transaction costs not modelled; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 7.0)

Why this score: The paper involves advanced reinforcement learning algorithms with formal risk constraints and mathematical formulations, while also providing extensive backtesting results, specific metrics, and an open-source repository.

Research Flowchart

  flowchart TD
  A["Research Goal: Create Risk-Sensitive Trading Agent"] --> B["Data Input:<br>Financial News (FNSPID) & Market Data (Nasdaq-100)"]
  B --> C["LLM Integration:<br>DeepSeek V3, Qwen 2.5, Llama 3.3<br>Generate Risk/Recommendation Signals"]
  C --> D["Core Methodology:<br>Risk-Sensitive CPPO (RL)<br>Conditional Value-at-Risk Optimization"]
  D --> E["Computation:<br>Backtesting & Agent Training"]
  E --> F["Key Findings:<br>Validated Risk-Sensitive Agents<br>Code & Models Available"]

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 risk-sensitive trading agent that combines reinforcement learning with large language models, extending CPPO with LLM-generated risk and recommendation signals.

Source sentence
We extend the Conditional Value-at-Risk Proximal Policy Optimization (CPPO) algorithm, by adding risk assessment and trading recommendation signals generated by a LLM from financial news.

Assumes 3

  • A 10% LLM infusion is used in the tests, meaning the original PPO and CPPO are modified up to 10%. model
    Source sentence
    In these tests, we use a 10% infusion of LLM, which means that the original PPO and CPPO up to 10%
  • One representative news article per stock per day is randomly selected, reducing the dataset to 2 million records. data
    Source sentence
    To reduce LLM API costs, we randomly select a representative news article per stock per day, reducing the dataset to 2 million records.
  • CVaR confidence level is set to 0.05, representing the worst 5% of outcomes. statistical
    Source sentence
    α is the CVaR confidence level (e.g., 0.05 for the worst 5%).

Data

Backtest sample: 2019 – 2023 · Nasdaq-100 index · equities · daily

  • FNSPID dataset · Dong et al., 2024 text-news unknown
    1999–2023 · Nasdaq-100 constituents
    Source sentence
    We use the FNSPID dataset (Dong et al., 2024), which contains 15.7 million time-aligned financial news records spanning 1999–2023.

Recognised sources: FNSPID

Costs

Not modeled A trading or execution result is reported without costs.

Finds 4

  • Integrating LLM stock recommendations consistently improves PPO cumulative returns, but does not yet beat the Nasdaq‑100 benchmark. performance
    Source sentence
    These preliminary results in figure 1 with Qwen 2.5 show that integrating LLM stock recommendations consistently improves PPO in terms of cumulative returns. However, it’s not sufficient to beat Nasdaq 100 benchmark so far.
  • PPO and CPPO‑DeepSeek outperform other methods, including the Nasdaq‑100 benchmark, in stochastic runs. performance
    Source sentence
    In these two runs (RL agents are stochastic), PPO and CPPO-DeepSeek outperform other methods, including the Nasdaq 100 benchmark.
  • Stronger LLM infusion generally degrades PPO‑DeepSeek performance, even with very small (0.1%) perturbations. negative
    Source sentence
    For PPO-DeepSeek, stronger LLM infusion mostly degrades performance, even for tiny (0.1%) perturbations.
  • Stronger LLM infusion improves CPPO‑DeepSeek performance. performance
    Source sentence
    For CPPO-DeepSeek, stronger LLM infusion improves performance

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 24 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 3 papers in the archive (3 by independent authors): Hierarchical Reinforced Trader (HRT): A Bi-Level Approach … (uses its method or data)

Code

https://github.com/benstaf/FinRL_DeepSeek — the authors' own repository

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