CLaC@FinMMEval 2026 Task 3: Sentiment-Augmented Deep Reinforcement Learning for Active Trading -- An Alpha-Reward Approach
Provides a detailed experimental framework and novel reward function for applying DRL to quantitative finance, directly relevant for researchers working on agent-based trading systems and generalization in RL.
AI Summary
This research paper details a system for the CLEF 2026 FinMMEval Lab trading task, comparing four Deep Reinforcement Learning algorithms (PG, PPO, DQL, DDPG) for making daily long/flat/short decisions on Bitcoin and Tesla. The system augments technical indicators and calendar encodings with market sentiment scores generated by LLaMA 3.2 1B and introduces an 'alpha reward' based on excess market return to reduce overfitting. Results show DDPG achieved the best overall test performance, with substantial validation-to-test generalization gaps highlighting the challenge of transferring policies across market regimes.
Excerpt
This paper presents our system for Task 3 of the CLEF 2026 FinMMEval Lab, which requires daily long, flat, or short trading decisions for Bitcoin (BTC) and Tesla (TSLA) using news and historical market data. We formulate the problem as a discrete-action Markov Decision Process and compare four deep reinforcement learning algorithms: Policy Gradient (PG), Proximal Policy Optimization (PPO), Deep Q-Learning (DQL), and Deep Deterministic Policy Gradient (DDPG). The agents use technical indicators,