Jul 2026· European Conference on Artificial Intelligence· pp. 1-12· 0 citations· 27 references
Abstract
This paper proposes an integrated framework for investment portfolio optimization that combines gradient boosting prediction models (LightGBM), volatility forecasting via the Heterogeneous Autoregressive Realized Volatility model (HAR-RV), market regime detection via Hidden Markov Models (HMM), and reinforcement learning agents. The system was implemented as an interactive web application (Streamlit) and evaluated on a diversified portfolio of 30 assets spanning six economic sectors, automatically selected from a universe of 530 tickers with a 40% sector concentration cap enforced at each rebalancing step to ensure structural diversification. Four reinforcement learning algorithms - PPO, SAC, TD3, and A2C - were evaluated under strictly identical experimental conditions to provide a comprehensive algorithmic comparison. The LightGBM return direction model achieved a walk-forward out-of-fold directional accuracy of 64.8%, while HAR-RV demonstrated consistent volatility forecasting superiority over LightGBM on 49 of 51 evaluated assets, motivating its adoption in the Black-Litterman-HRP covariance adjustment. Over the full 2015-2024 period, SAC achieved the highest Sharpe ratio (1.193) and Jensen's Alpha (+10.58% annual), followed by PPO (Sharpe 1.157). Out-of-sample (2021-2024), SAC maintained dominance (Sharpe 1.002, Alpha +6.33%), confirmed robust across eight distinct evaluation windows in a temporal sensitivity analysis. Deflated Sharpe Ratio analysis (p_DSR = 1.000 for all strategies) highlights the inherent statistical limitation of ~2.5-year evaluation horizons, motivating the presented framework as a validated software architecture for integrated ML+RL portfolio management rather than a definitively proven superior strategy.
This study applies deep reinforcement learning (DRL) to multi-asset portfolio optimization in the Vietnamese stock market, aiming to evaluate the performance and stability of different DRL algorithms under emerging market conditions.
Seven algorithms – A2C, proximal policy optimization (PPO), deep deterministic policy gradient, twin delayed deep deterministic policy gradient (TD3), soft actor-critic (SAC), truncated quantile critics (TQC) and RecurrentPPO – are trained on daily data from January 2018 to December 2024 and evaluated out-of-sample from January 2023 through September 2025 using five liquid equities (SBT.VN, BID.VN, CTG.VN, HPG.VN and VCB.VN). The state representation includes technical indicators (RSI, MACD, SMA, EMA, Bollinger Bands and OBV), as well as risk features such as rolling covariance and a turbulence index.
PPO achieves the highest annual return (0.1666) and demonstrates the most stable performance. TD3 delivers comparable cumulative growth with higher variability. RecurrentPPO attains the highest Sharpe ratio (1.0461), highlighting the importance of temporal modeling. SAC and TQC produce more conservative but stable outcomes.
This study does not propose a new DRL algorithm; instead, it provides a controlled and reproducible benchmarking framework for comparing multiple DRL models under identical conditions in an emerging market setting.
W. Saijai, Kansuda Pankwaen· Asian Journal of Economics a...· 0 citations
This paper explores the effectiveness of integrating Artificial Neural Network (ANN) based return forecasts into three portfolio optimization frameworks: Equal-Weight (EW), Mean-Variance (MV), and Black-Litterman (BL), under varying market regimes, using 67 stocks listed on Thailand’s Market for Alternative Investment (MAI) as a case study. Portfolios are constructed using ANN predictions and tested across pre-COVID stable conditions (2019) and a volatile period (2020–2024), with an additional evaluation of rebalanced portfolios on 2024 data. Results indicate that BL achieves the highest risk-adjusted returns under stable conditions, while EW outperforms significantly during high volatility, driven largely by outlier stock performance during the COVID-19 recovery. Rebalancing with updated ANN forecasts did not guarantee improved performance, highlighting both the promise and limitations of machine learning-enhanced optimization in emerging markets.
Pema Yangchen, Rujira Chaysiri· International Conference on...· 0 citations
This paper introduces a dynamic portfolio optimization framework for large institutional investors using Scientific Physics-Informed Reinforcement Learning (SciPhyRL). Formulated in continuous time over an extended state space that includes explicit cumulative costs, the approach leverages offline historical data to learn optimal, distribution-aware strategies. A core innovation reduces the optimization challenge to solving an HJB equation by projecting it onto observed trajectories as a pathwise Hamilton-Jacobi equation. This is solved directly from data using PINN in a single offline sweep, eliminating the need for traditional value or policy iteration. To make the method effective at practical short horizons, the control variable is recast from a continuous trading rate to a discrete target holding. This ensures signal-implied positions are reached immediately, while execution costs are evaluated against a microstructure-grounded quadratic price impact model. Evaluated on a $14$-asset ETF universe using an engineered oracle signal, the learned Gibbs policy yields substantial out-of-sample Sharpe ratio improvements over static and myopic baselines. The results demonstrate that the proposed framework successfully translates known signal quality into a robust, multi-period, and cost-aware allocation mechanism with strictly controlled volatility and turnover.
This paper asks which families of volatility forecasting models create economic value in active trading, and through which integration channel that value is transmitted. Seven models drawn from four families — econometric (GJR-GARCH, HAR-J), gradient boosting (XGBoost, LightGBM), recurrent neural networks (LSTM, GRU) and a hybrid combining HAR-J with boosting — are compared on the realized volatility of 17 liquid Moscow Exchange stocks computed from 10-minute returns over 2014–2026, using a feature space of 234 variables and QLIKE as the loss function. Two findings follow. First, the machine learning advantage is neither uniform across families nor stable across horizons: gradient boosting improves on the HAR-J benchmark by 5–9% at the one-day horizon but loses to it at the five-day horizon under rolling re-estimation, whereas recurrent networks are 22–31% worse than HAR-J at the one-day horizon and dominate it at no horizon; only the hybrid ranks at or near the top throughout. Second, the translation of accuracy into trading performance is governed by the integration channel rather than by the size of the accuracy gain: a 28% reduction in QLIKE raises the Sharpe ratio of a six-strategy portfolio by +0.79 and +0.85 under regime filtering and volatility targeting (p < 0.001), but by an insignificant +0.04 when the forecast only scales the parameters of an individual trade. Model selection therefore pays where the forecast governs the entry decision and the position size, and not otherwise. Statistical accuracy is a valid proxy for economic value in this setting, but only conditionally on the channel through which the forecast reaches the trading decision.
Nikita I. Lysenok· International journal of com...· 0 citations
Despite the dynamic nature of the market, large dimensionalities of asset space and varieties of financial products, portfolio optimization is a highly challenging problem. These traditional methods like Markowitz Mean-Variance Optimization and Risk Parity are highly static, and have not been able to adjust to the speed-of-change in market regimes. In this paper, the authors present a new Quantum-Inspired Portfolio Optimization (QIPO-RL) model that combines these interesting approaches to create a framework for a Reinforcement Learning (RL) based dynamic stock allocation algorithm. The framework blends quantum-inspired search techniques, an adaptive RL agent, and asset weights to maximize risk-adjusted asset returns, while maintaining assets in optimal allocation, and provides a way to rebalance a portfolio continuously in response to the changing market environment. The results from experiments were compared with state of the art baselines, which showed that QIPO-RL’s annual return is 16.8%, its Sharpe ratio is 1.61, and the maximum drawdown is 11.5% which is the best among all the competing methods. These findings support the synergism of using a global search method inspired by quantum computers in conjunction with an RL-based adaptation of decisions.
Kishore Kumar Sambangi· International Journal of Com...· 0 citations