CLaC@FinMMEval 2026 Task 3: Sentiment-Augmented Deep Reinforcement Learning for Active Trading -- An Alpha-Reward Approach

2026-07-17 · Source: Machine Learning · Field: Finance & Economics — Capital Markets & Investment Management, FinTech & Digital Financial Services · Depth: Expert, quick

Summary

The CLaC@FinMMEval 2026 Task 3 system addresses daily active trading decisions for Bitcoin (BTC) and Tesla (TSLA) by formulating the problem as a discrete-action Markov Decision Process. It compares four deep reinforcement learning algorithms: Policy Gradient (PG), Proximal Policy Optimization (PPO), Deep Q-Learning (DQL), and Deep Deterministic Policy Gradient (DDPG). Agents integrate technical indicators, cyclical calendar encodings, and daily news sentiment scores from LLaMA 3.2 1B. To mitigate overfitting and align with outperforming buy-and-hold, an alpha reward based on excess market return and randomized episode start dates were introduced. Hyperparameters were optimized using Ray Tune across 180 trials per algorithm-asset pair, with DDPG achieving the strongest overall performance on the CLEF Task 3 test set. For TSLA, DDPG and DQL yielded cumulative returns of 54.96% and 52.62% respectively, significantly surpassing buy-and-hold's 16.45%. DDPG also generated a positive 1.58% return for BTC, while buy-and-hold declined by -34.27%. The study noted a significant generalization gap between validation and test sets, indicating challenges in transferring policies from bull to bear market conditions.

Key takeaway

For Machine Learning Engineers developing active trading strategies, this research indicates that sentiment-augmented deep reinforcement learning, particularly DDPG or DQL, can deliver significant returns over traditional buy-and-hold. Your implementation should incorporate LLaMA 3.2 1B for news sentiment and utilize alpha rewards with randomized episode starts to enhance robustness. However, be vigilant about the substantial validation-to-test generalization gap, especially when transferring policies between different market regimes, and ensure your models are rigorously tested across varied market conditions.

Key insights

Sentiment-augmented deep reinforcement learning can significantly outperform buy-and-hold in active trading, despite generalization challenges across market regimes.

Principles

• Alpha rewards based on excess market return align DRL with outperforming buy-and-hold.
• Randomizing episode start dates helps reduce overfitting in DRL trading.
• Policy transfer from bull to bear markets faces a substantial generalization gap.

Method

Formulate active trading as a discrete-action Markov Decision Process, integrating technical indicators, calendar encodings, and LLM-derived news sentiment, with hyperparameter optimization via validation Sharpe ratio.

In practice

• Employ DDPG or DQL for active trading strategies.
• Integrate LLaMA 3.2 1B for news sentiment in trading.
• Utilize alpha rewards and episode randomization for DRL robustness.

Topics

• Deep Reinforcement Learning
• Active Trading
• Financial Sentiment Analysis
• LLaMA 3.2 1B
• Bitcoin Trading
• Tesla Trading
• Sharpe Ratio

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Dynamic Multi-Pair Trading Strategy in Cryptocurrency Markets with Deep Reinforcement Learning — DRL execution overlays, shielded by statistical boundaries, significantly enhance crypto pair trading performance and safety.
• Evaluating Large Language Model News Sentiment in Finance under Liquidity and Market Frictions — High-capacity LLMs extract economically meaningful financial news sentiment even under realistic market frictions, unlike simpler lexicon methods.
• Codex 5.5 vs Claude Opus 4.7 Polymarket Trading Challenge — Specific AI trading strategies can outperform rivals in short-term, high-frequency markets by exploiting market inefficiencies.
• Introduction to quantitative finance Part 26: Whether to use forecasting methods, or to tell an LLM… — Forecasting in quantitative finance faces challenges with LLMs, necessitating memory-enhanced reward shaping for consistency.
• AI stock trading startup EquiLibre raises funding at a $500M+ valuation — Self-play AI, proven in imperfect-information games, is effectively applied to high-stakes financial trading.
• CSTrader: A Testbed for Language-Grounded Trading in a Community-Driven Virtual Asset Market — Niche, language-driven markets offer a robust benchmark for language-to-action LLM research.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.