Paper: arXiv 2412.17293

Authors: Sumit Nawathe, Ravi Panguluri, James Zhang, Sashwat Venkatesh

Abstract

We propose a reinforcement learning (RL) framework that leverages multimodal data including historical stock prices, sentiment analysis, and topic embeddings from news articles, to optimize trading strategies for SP100 stocks. Building upon recent advancements in financial reinforcement learning, we aim to enhance the state space representation by integrating financial sentiment data from SEC filings and news headlines and refining the reward function to better align with portfolio performance metrics. Our methodology includes deep reinforcement learning with state tensors comprising price data, sentiment scores, and news embeddings, processed through advanced feature extraction models like CNNs and RNNs. By benchmarking against traditional portfolio optimization techniques and advanced strategies, we demonstrate the efficacy of our approach in delivering superior portfolio performance. Empirical results showcase the potential of our agent to outperform standard benchmarks, especially when utilizing combined data sources under profit-based reward functions.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • Empirical Rigor: 6.0/10 (capped from 8.0: transaction costs only partially 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 employs advanced mathematical concepts including Markov Decision Processes, deep neural networks, and tensor-based feature extraction, while also detailing comprehensive data sourcing, preprocessing pipelines, and empirical benchmarking against traditional methods.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Optimize SP100 Trading via RL<br>with Multimodal Data"] --> B["Data Inputs & Processing"]
  subgraph B ["Inputs"]
      B1["Price History"]
      B2["News Sentiment"]
      B3["SEC Filings"]
  end
  B --> C["Deep RL Agent Construction"]
  C --> D["Feature Extraction:<br>CNN/RNN on State Tensors"]
  D --> E["Policy Learning:<br>Profit-based Reward Function"]
  E --> F["Execution & Backtesting"]
  F --> G["Key Findings:<br>Outperforms Benchmarks<br>Synergy of Data Sources"]

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 reinforcement learning framework that uses multimodal data—historical prices, sentiment, and news embeddings—to optimize trading strategies for S&P 100 stocks.

Source sentence
We propose a reinforcement learning (RL) framework that leverages multimodal data—including historical stock prices, sentiment analysis, and topic embeddings from news articles—to optimize trading strategies for S&P100 stocks.

Assumes 4

  • There are m tradable stocks in the universe and one risk‑free asset, and portfolio weights sum to one. model
    Source sentence
    Assume that there are m tradable stocks in our universe and one risk-free asset. The action space A = a ∈Rm+1 |
  • State transitions are deterministic (delta function). model
    Source sentence
    The transition function T is a delta function since state transitions are deterministic.
  • Sentiment scores decay exponentially over days since the last report. data
    Source sentence
    where a represents the company’s sentiment score on the most recent reporting date, t represents the time (in days) between the last report date and the current day, and γ is a constant between 0 and 1 representing the daily decay factor.
  • The reward function can be pure profit or a differential Sharpe ratio. model
    Source sentence
    One simple reward function is pure profit: R(st, at) is how much profit is gained

Data

Backtest sample: S&P100 stocks · equities · daily

  • NewsHeadlines dataset text-news unknown
    84 S&P 100 tickers
    Source sentence
    Our dataset contains data for 84 out of the total 100 tickers in the S&P 100, and it contains 70,872 entries containing the sentiment embedding of news for a company on a given day.
  • Historical stock price data market-prices unknown
    S&P 100 stocks
    Source sentence
    We propose a reinforcement learning (RL) framework that leverages multimodal data—including historical stock prices, sentiment analysis, and topic embeddings from news articles—to optimize trading strategies for S&P100 stocks.

Recognised sources: CRSP WRDS SEC EDGAR Kaggle

Costs

Partial Costs appear only as a robustness check, a flat fee without spread or impact, or outside the headline numbers.

Finds 4

Reported performance (SEC+News CNN (Profit)): Sharpe 0.55 (benchmark 0.55) · total return 16.6% · max drawdown 19.5%

  • Agents trained with the profit reward consistently achieve superior portfolio performance compared to those using the differential Sharpe ratio. performance
    Source sentence
    Our principal observations indicate a disparity in learning complexity between the profit reward function and the differential Sharpe ratio for our agent, resulting in consistently superior portfolio performance when optimizing with the former.
  • CNN‑EIIE and RNN‑EIIE policies outperform MLP under the profit reward, while MLP tends to overfit. performance
    Source sentence
    agents trained on CNN EIIE and RNN EIIE, exhibit enhanced performance under the profit reward, while the MLP policy network seems to overfit significantly.
  • Integrating SEC filings data improves performance across both reward functions relative to baseline models. performance
    Source sentence
    Upon integrating SEC filings data, notable performance enhancements are observed across both reward functions compared to baseline models.
  • Combining news and SEC data yields the best performance, but news data alone hampers learning of the differential Sharpe ratio due to sparsity. performance
    Source sentence
    Optimal performance is achieved when simultaneously integrating news and SEC data into the agent’s environment.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 32 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 1 paper in the archive (1 by independent authors)

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 (15)

  1. MAPS : Multi -agent Reinforcement Learning -based Portfolio Management System (2020) — link
  2. Stock portfolio selection using learning-to-rank algorithms with news sentiment (2017) — link
  3. A Deep Reinforcement Learning Framework for the Financial Portfolio Management Problem (2017) — link
  4. Reinforcement- Learning based Portfolio Management with Augmented Asset Movement Prediction States (2020) — link
  5. Deep reinforcement learning for stock portfolio optimization by connecting with modern portfolio theory (2023) — link
  6. Deep Reinforcement Learning for Optimal Portfolio Allocation : A Comparative Study with Mean - Variance Optimization — → 2602.17098
  7. FinBERT - FOMC : Fine - Tuned FinBERT Model with Sentiment Focus Method for Enhancing Sentiment Analysis of FOMC Minutes (2023) — link
  8. Reinforcement Learning for Trading Systems and Portfolios: Immediate vs Future Rewards — link
and 7 more
  1. Optimization of trading systems and portfolios — link
  2. On-Line Portfolio Selection with Moving Average Reversion — link
  3. Weighted Moving Average Passive Aggressive Algorithm for Online Portfolio Selection — link
  4. Daily Finaincial News for 6000+ Stocks (2020)
  5. CRSP Daily Stocks (2010-2024) — link
  6. SEC .gov \textbar How to Read a 10-K/10-Q — link
  7. SEC .gov \textbar EDGAR \textbar Company Filings — link

1 of 15 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.