Meta-Optimized Risk-Aware Portfolio Management: A Hybrid Deep Reinforcement Learning and LSTM-GRU Ensemble

Majid Sorouri; Deniz NoorMohammadzadehMaleki; Arman Salehi; Amirfarhad Farhadi; Azadeh Zamanifar

doi:10.1109/SEEDA-CECNSM68644.2025.11329752

Research question

How can forecasting, policy learning, and downside-risk control be coordinated for sequential portfolio allocation?

Condensed abstract

The paper presents a hybrid architecture combining an LSTM–GRU forecasting ensemble, deep reinforcement-learning agents, and meta-optimization for risk-aware portfolio decisions.

Key contribution

A forecasting-and-control architecture centered on explicit downside-risk objectives and adaptive optimization.

Method overview

Sequence forecasting, deep reinforcement learning, a CVaR-oriented objective, and meta-optimization are combined in a portfolio-management pipeline.

The average hides the tail

Two distributions. The same average loss. Switch between them to see why a risk-aware objective pays attention to what happens at the extreme.

Ordered losses · arbitrary units Selected worst-case tail

Mean loss13.2Unchanged in both presets

VaR at 80%19.0Empirical loss quantile

Illustrative CVaR29.52 of 10 scenarios

Worst-case tail 20%Risk penalty λ 0.50

Illustrative objective · lower is preferred

mean loss + λ × CVaR

13.20 + 14.7527.95

The worst 20% contains losses 25 and 34. Their average is 29.5 loss units. The mean across all ten scenarios stays 13.2 in both distributions.

See all ten synthetic losses

Scenario 12Scenario 24Scenario 35Scenario 47Scenario 59Scenario 612Scenario 715Scenario 819Scenario 925Scenario 1034

Each scenario has probability 0.1. VaR uses the empirical left quantile; CVaR averages the selected upper-tail scenarios. Tail sizes are whole multiples of one scenario.

Where tail risk enters the architecture

An LSTM–GRU ensemble supplies sequence-based forecasts. A forecast is an input to a decision, not the decision itself.

Synthetic losses in arbitrary units, not market returns, backtest results, or investment advice. The loss-plus-CVaR expression is an educational objective, not a reproduction of the paper’s trained policy. Published forecasting-and-control architecture.

Dataset and evaluation

Data

Financial-market data as described in the conference paper. No public data package is linked from this portfolio.

Protocol

Backtesting and Bayesian analysis are reported in the proceedings paper. The portfolio does not add experimental detail that cannot be checked against the full record.

Main results

The proceedings paper reports quantitative backtesting results; this site intentionally does not restate point estimates that have not been independently verified from the full paper.

Baseline comparison

No broad superiority claim is made. Conclusions remain conditional on the paper’s market period, cost assumptions, baselines, and backtesting protocol.

Limitations and failure modes

Financial backtests are sensitive to market regime, transaction costs, leakage, and selection effects.
No release-grade reproduction package is linked from the public portfolio.
The work is methodological and must not be interpreted as investment advice.

Reproducibility

A synthetic loss-tail calculation that makes the risk objective inspectable.

Teaching example onlyChecked locally 2026-09-05 · bundle v1.0.0

Paper-associated code: No public full-paper code archive is linked by this portfolio.
Inputs for this check: Two invented ten-scenario loss distributions, a 20% upper tail, and penalty λ = 0.5.
Prerequisites: Python 3.10+ · standard library only · CPU · no network or large dataset download. A notebook environment is optional.

Run the small check

Download and extract the complete bundle.
Open a terminal in the extracted folder and run python check.py.
Calculate mean loss, empirical left-quantile VaR, upper-tail CVaR, and the illustrative loss-plus-CVaR objective.

Download check bundle Read the guide

Includes Python, an equivalent notebook, source JSON, expected output, and SHA-256 provenance. The notebook runs from the extracted folder.

Expected output and individual files

Equal means: 13.2. CVaR: 29.5 versus 45.5. Objectives: 27.95 versus 35.95.

Python script Notebook Input data Expected JSON Provenance

What this does not establish

No public full-paper code package is linked here. This example does not reproduce forecasts, a trained allocation policy, market returns, or a backtest; it is not investment advice.

These checks use the portfolio’s existing sourced records. External archive links could not be freshly rechecked during package preparation. A passing check is not a full-paper replication.

The decisions behind the method

Why sequential decision systems need explicit downside objectives and skeptical backtesting.

Read the research note

Verification sources

Cite this work

Download .bib Download .ris

Majid Sorouri, Deniz NoorMohammadzadehMaleki, Arman Salehi, Amirfarhad Farhadi, Azadeh Zamanifar (2025). Meta-Optimized Risk-Aware Portfolio Management: A Hybrid Deep Reinforcement Learning and LSTM-GRU Ensemble. 2025 10th South-East Europe Design Automation, Computer Engineering, Computer Networks and Social Media Conference (SEEDA-CECNSM), 1–6. https://doi.org/10.1109/SEEDA-CECNSM68644.2025.11329752

BibTeX

@inproceedings{sorouri2025riskaware,
  title = {Meta-Optimized Risk-Aware Portfolio Management: A Hybrid Deep Reinforcement Learning and LSTM-GRU Ensemble},
  author = {Majid Sorouri and Deniz NoorMohammadzadehMaleki and Arman Salehi and Amirfarhad Farhadi and Azadeh Zamanifar},
  year = {2025},
  doi = {10.1109/SEEDA-CECNSM68644.2025.11329752},
  booktitle = {2025 10th South-East Europe Design Automation, Computer Engineering, Computer Networks and Social Media Conference (SEEDA-CECNSM)},
  pages = {1--6},
  publisher = {IEEE}
}

Publication record last verified 2026-07-26.

Discuss this research

Want to discuss tail estimation, sensitivity to the risk penalty, or evaluation under distribution shift? Include a synthetic setting to make the question concrete.

Draft a research question

Opens your email app with the paper and current experiment. Nothing is sent automatically.