Paper: arXiv 2509.24144

Authors: Yun Lin, Jiawei Lou, Jinghe Zhang

Abstract

Deep learning offers new tools for portfolio optimization. We present an end-to-end framework that directly learns portfolio weights by combining Long Short-Term Memory (LSTM) networks to model temporal patterns, Graph Attention Networks (GAT) to capture evolving inter-stock relationships, and sentiment analysis of financial news to reflect market psychology. Unlike prior approaches, our model unifies these elements in a single pipeline that produces daily allocations. It avoids the traditional two-step process of forecasting asset returns and then applying mean–variance optimization (MVO), a sequence that can introduce instability. We evaluate the framework on nine U.S. stocks spanning six sectors, chosen to balance sector diversity and news coverage. In this setting, the model delivers higher cumulative returns and Sharpe ratios than equal-weighted and CAPM-based MVO benchmarks. Although the stock universe is limited, the results underscore the value of integrating price, relational, and sentiment signals for portfolio management and suggest promising directions for scaling the approach to larger, more diverse asset sets.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 3.0/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper employs advanced deep learning architectures (LSTMs, GATs) with mathematical formulations and a custom Sharpe-ratio loss function, but its empirical validation is limited to a small universe of nine stocks without robust out-of-sample testing or data availability details.

Research Flowchart

  flowchart TD
  A["Research Goal: End-to-End<br>Portfolio Optimization"] --> B["Data Inputs:<br>Price, News, Relations"]
  B --> C["Methodology Pipeline:<br>LSTM + GAT + Sentiment"]
  C --> D["Computational Process:<br>Direct Weight Generation"]
  D --> E["Key Findings:<br>Higher Returns & Sharpe Ratio"]
  E --> F["Outcome:<br>Proof of Concept for<br>Integrated Signal Processing"]

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 an end‑to‑end deep learning framework that directly learns portfolio weights by combining LSTM temporal modeling, GAT relational modeling, and sentiment analysis of financial news.

Source sentence
We present an end-to-end framework that directly learns portfolio weights by combining Long Short-Term Memory (LSTM) networks to model temporal patterns, Graph Attention Networks (GAT) to capture evolving inter-stock relationships, and sentiment analysis of financial news to reflect market psychology.

Assumes 5

  • The investable universe is restricted to the nine selected stocks, with no rebalancing into new assets. model
    Source sentence
    The investable universe is restricted to the nine selected stocks, with no rebalancing into new assets.
  • Trades are assumed to execute perfectly at official open or close prices, with zero slippage or delay. execution
    Source sentence
    Trades are assumed to execute perfectly at official open or close prices, with zero slippage or delay.
  • All trading data are treated as correctly timestamped and free of missing values. data
    Source sentence
    All trading data are treated as correctly timestamped and free of missing values.
  • Sector membership is assumed constant, ignoring possible corporate restructuring. model
    Source sentence
    Sector membership is assumed constant, ignoring possible corporate restructuring.
Show 1 more
  • Transaction costs, bid‑ask spreads, and market impacts are excluded. execution
    Source sentence
    Finally, transaction costs, bid-ask spreads, and market impacts are excluded.

Data

Backtest sample: 2024-01 – 2025-06 · nine US stocks across six sectors · equities · daily

  • AlphaVantage daily price data · AlphaVantage market-prices commercial
    January 2021 to May 2025 · nine U.S. stocks across six sectors
    Source sentence
    Daily price data (open, high, low, close, and volume) were obtained from the AlphaVantage API [1] for the period January 2021 to May 2025
  • MarketAux news data · MarketAux text-news commercial
    January 2021 to May 2025 · nine U.S. stocks
    Source sentence
    News data were retrieved from the MarketAux API [9], which provides timestamped articles with metadata including title, snippet, relevance score, and sentiment score
  • S&P 500 index benchmark · yfinance benchmark public
    S&P 500
    Source sentence
    Benchmark data were also collected: S&P 500 index levels (via the yfinance API)
  • 3‑month U.S. Treasury yields · FRED macro public
    Source sentence
    3-month U.S. Treasury yields (via FRED [4]) to construct CAPM-MVO baselines

Recognised sources: Alpha Vantage

Costs

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

Source sentence
Finally, transaction costs, bid-ask spreads, and market impacts are excluded.

Finds 2

Reported performance (LSTM‑GAT end‑to‑end portfolio model (Model v3)): Sharpe 1.15 (benchmark 0.83) · annual return 31.2% · total return 37.3% · max drawdown 22.1%

  • The model delivers higher cumulative returns and Sharpe ratios than equal‑weighted and CAPM‑based MVO benchmarks. performance
    Source sentence
    In this setting, the model delivers higher cumulative returns and Sharpe ratios than equal-weighted and CAPM-based MVO benchmarks.
  • Sentiment features improved returns and Sharpe ratios, and PCA helped stabilize allocations and limit drawdowns. performance
    Source sentence
    Sentiment features improved returns and Sharpe ratios, while dimensionality reduction via PCA helped stabilize allocations and limit drawdowns.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 34 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 (16)

  1. Deep learning for portfolio optimization (2020) — link
  2. Large-scale time-varying portfolio optimisation using graph attention networks — → 2407.15532
  3. Graph neural networks for asset management (2021) — link
  4. Graph neural networks for deep portfolio optimization (2023) — link
  5. HGFormer: A hierarchical graph transformer for portfolio optimization (2024) — link
  6. Deep learning in finance (2016) — link
  7. Leveraging BiLSTM-GAT for enhanced stock market prediction: a dual-graph approach to portfolio optimization (2025) — link
  8. Effects of Daily News Sentiment on Stock Price Forecasting — → 2308.08549
and 8 more
  1. Long Short-Term Memory (1997) — link
  2. Two‑Stage Portfolio Optimization Integrating Optimal Sharpe Ratio Measure and Ensemble Learning (2022) — link
  3. High-performance stock index trading: making effective use of a deep long short-term memory neural network (2019) — link
  4. Graph Attention Networks (2018) — link
  5. MarketAux API (2025)
  6. Alpha Vantage API (2025)
  7. 3-Month Treasury Bill: Secondary Market Rate (TB3MS) (2025)
  8. Adam: A Method for Stochastic Optimization (2014) — link

2 of 16 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.