Hybrid DDQN improves trading-action classification using sentiment and financial signals

A sentiment-enhanced double deep Q-network produced the strongest reported trading-action classification results in a computational study that combined market sentiment with conventional financial indicators. The hybrid model achieved an accuracy of 0.9556, precision of 0.9605, recall of 0.9556 and F1-score of 0.9558 when classifying hold, invest and sell actions. It outperformed four comparison approaches—BiLSTM with attention, proximal policy optimization, a conventional long short-term memory network and a standard deep Q-network—on the performance metrics summarized in the accessible article abstract.

The research treats portfolio optimization as a reinforcement-learning problem in which an algorithm associates changing market conditions with trading actions and rewards. The authors say sequence models such as LSTM and BiLSTM with attention can identify temporal relationships and sequences in financial data, but do not directly incorporate shifts in investor psychology or detect rapid changes in sentiment. They also identify limitations in reinforcement-learning systems based on DQN and proximal policy optimization, including poor reward propagation, weak exploration in unstable environments and difficulty using several information sources. Their proposed double deep Q-network, or DDQN, was designed to address these issues by incorporating both sentiment-derived and engineered financial signals into the decision process.

The two types of input provide complementary information. A sentiment-analysis component processes signals arising from news and social media to represent investor sentiment and associated volatility, while engineered financial features capture price, volume and momentum information. The abstract describes the sentiment module as capable of analyzing market sentiment and volatility and the financial module as supplying conventional market indicators. Together, these inputs are intended to let the strategy respond to psychological and event-related dynamics as well as patterns in financial data. The authors also report reward-per-stock results suggesting that the algorithm can operate across both stable and volatile stocks under the study’s experimental conditions, with cumulative rewards corresponding to Sharpe ratios of more than approximately 3.0–6.0.

Classification analysis showed that the hybrid model made the fewest errors among the hold, invest and sell categories. Receiver operating characteristic curves further indicated that these actions were most separable under the model, with areas under the curve of 0.94 for hold, 0.92 for invest and 0.90 for sell. Its F1-score of 0.9558 was higher than the reported values of 0.796 for BiLSTM with attention, 0.781 for proximal policy optimization, 0.767 for LSTM and 0.754 for DQN; corresponding accuracies were 0.795, 0.780, 0.768 and 0.755. These are computational results from the reported experimental comparison, not demonstrated live-trading returns. The accessible abstract does not specify the tested period, market, asset universe, sample size, transaction-cost assumptions, human-investor comparison or prospective performance, so the results do not establish that the framework would consistently reproduce its reported accuracy or Sharpe ratios in real markets.

Subject of Research: Evaluation of a sentiment-enhanced deep reinforcement learning framework for stock-market forecasting and portfolio optimization

Article Title: Sentiment-enhanced deep reinforcement learning framework for stock market forecasting and portfolio optimization

Article References: Altaf, H., & Yadav, P. K. (2026). Sentiment-enhanced deep reinforcement learning framework for stock market forecasting and portfolio optimization. International Journal of Data Science and Analytics, 22(1), Article 348. https://doi.org/10.1007/s41060-026-01339-8

Image Credits: AI Generated

DOI: 10.1007/s41060-026-01339-8

Keywords: deep reinforcement learning, double deep Q-network, market sentiment, stock forecasting, portfolio optimization, trading actions, financial signals, Sharpe ratio, Sentiment-enhanced, deep, reinforcement, learning