Paper: arXiv 2309.16679

Authors: Paraskevi Nousi, Loukia Avramelou, Georgios Rodinos, Maria Tzelepi, Theodoros Manousis, Konstantinos Tsampazis, Kyriakos Stefanidis, Dimitris Spanos, Manos Kirtas, Pavlos Tosidis, Avraam Tsantekidis, Nikolaos Passalis, Anastasios Tefas

Abstract

Financial portfolio management describes the task of distributing funds and conducting trading operations on a set of financial assets, such as stocks, index funds, foreign exchange or cryptocurrencies, aiming to maximize the profit while minimizing the loss incurred by said operations. Deep Learning (DL) methods have been consistently excelling at various tasks and automated financial trading is one of the most complex one of those. This paper aims to provide insight into various DL methods for financial trading, under both the supervised and reinforcement learning schemes. At the same time, taking into consideration sentiment information regarding the traded assets, we discuss and demonstrate their usefulness through corresponding research studies. Finally, we discuss commonly found problems in training such financial agents and equip the reader with the necessary knowledge to avoid these problems and apply the discussed methods in practice.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • Empirical Rigor: 4.0/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper covers advanced deep learning architectures (RNNs, LSTMs, Transformers, DRL) and discusses mathematical concepts like non-stationarity and normalization, but the excerpt focuses on literature review and theoretical challenges rather than presenting specific backtests, code, or statistical performance metrics.

Research Flowchart

  flowchart TD
  A["Research Goal: DL for Financial Portfolio Management"] --> B["Data Collection & Preprocessing"]
  
  subgraph B ["Data Sources"]
      B1["Price Data<br>OHLCV"]
      B2["Sentiment Data<br>Online Sources"]
  end

  B --> C["Methodology"]
  
  subgraph C ["Dual Approach"]
      C1["Supervised Learning<br>for Prediction"]
      C2["Reinforcement Learning<br>for Decision Making"]
  end

  C --> D["Computational Process<br>Deep Learning Models"]
  
  subgraph D ["Model Types"]
      D1["Recurrent Networks<br>LSTM/GRU"]
      D2["Transformers<br>Attention Mechanism"]
      D3["Actor-Critic RL<br>Agent"]
  end

  D --> E["Key Findings & Outcomes"]
  
  subgraph E ["Results"]
      E1["Sentiment Enhances<br>Performance"]
      E2["RL Outperforms<br>Traditional Methods"]
      E3["Multi-Asset<br>Optimization"]
      E4["Practical Training<br>Guidelines"]
  end

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 adaptive data normalization methods, reward shaping policies, and the integration of sentiment information to improve deep learning based financial trading agents.

Source sentence
... the proposed reward shaping policy proved helpful in both cases compared to the baseline PnL and Fee rewards.

Assumes 5

  • Commission fee c is set slightly higher than the actual commission to account for price slippage. execution
    Source sentence
    c is typically set to a value slightly higher than the actual commission paid to account for price slippage and possible risks.
  • Financial time‑series measurements are generated from a Gaussian (or Gaussian mixture) distribution for normalization. statistical
    Source sentence
    Assuming a Gaussian distribution for each measurement vector
  • A threshold (pthres) is set high enough to overcome any commission fees and price slippage. execution
    Source sentence
    Typically, pthres is set to a value high enough to overcome any commission fees, as well as to account for price slippage that might occur.
  • Agents can exit the market without terminating training and may re‑enter later. execution
    Source sentence
    Note that usually exiting the market does not terminate the training process and the agent has the opportunity to reenter in it.
Show 1 more
  • Sentiment extracted from online sources can be a stronger predictor than price alone for Bitcoin trading. other
    Source sentence
    demonstrating that sentiment information can be a stronger predictor for Bitcoin trading compared to the actual price

Data

  • FI-2010 order-book public
    5 NASDAQ stocks over 10 consecutive days
    Source sentence
    ... contains 4 million time series data for five stocks from Nasqad Nordic market in a time frame of 10 consecutive days.
  • Bitcoin and Ethereum price data crypto-onchain public
    Cryptocurrencies Bitcoin and Etherium
    Source sentence
    on ... the popular cryptocurrencies BitCoin and Etherium.
  • Financial dataset with 28 instrument combinations market-prices unknown
    2009 to mid‑2018 · 28 different instrument combinations with currencies such as Euro, Dollar, British pounds
    Source sentence
    evaluated on ... a financial dataset that contains 28 different instrument combinations with currencies such as Euro, Dollar, British pounds and Canadian dollar among others. The dataset contains minute price candles starting from 2009 to mid-2018.

Recognised sources: FI-2010

Costs

Modeled Costs are part of the reported headline results.

Commission fee c is applied each step as ct = |at−at−1|c and is set slightly higher than actual to cover slippage.

Source sentence
... c is typically set to a value slightly higher than the actual commission paid to account for price slippage and possible risks.

Finds 3

  • PPO with the proposed price‑trailing method improves PnL from 23.5% ± 2.0% to 28.6% ± 2.4% and Sharpe ratio from 3.59 ± 0.37 to 4.20 ± 0.43. performance
    Source sentence
    PnL 23.5% ± 2.0% → 28.6% ± 2.4% ... Sharpe ratio 3.59 ± 0.37 → 4.20 ± 0.43
  • Adaptive normalization (RDAIN) outperforms static and other adaptive methods on FI‑2010 in terms of F1‑score and Cohen’s κ. performance
    Source sentence
    it is evident that the proposed method outperforms all of the compared normalization methods, including an adaptive method (DAIN) lacking the gating mechanism
  • Combining price with sentiment yields higher returns (10.55 ± 1.16) and Sharpe (156.68 ± 5.38) than price only (9.68 ± 4.93, 149.08 ± 5.37). performance
    Source sentence
    Price only 9.68 ± 4.93 149.08 ± 5.37 ... Price + Sentiment 10.55 ± 1.16 156.68 ± 5.38

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 19 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.