Mining Evidences for Concept Stock Recommendation.

Qi Liu, Yue ZhangNAACL 2018 · DOI · OpenReview

TL;DR

This paper tackles the challenge of recommending relevant stocks to investors based on abstract "concepts" (like economic policies or events), especially in emerging markets with limited structured data. They propose a novel reinforcement learning approach that iteratively expands these concepts by mining textual evidence, outperforming traditional methods.

What Problem Does This Paper Address?

Identifying stocks related to specific "concepts" (e.g., a new economic region like Xiong'an or a legal ruling on intellectual property) is crucial for investors. However, this is difficult because:

  1. Lack of Structural Understanding: The relationship between concepts and stocks isn't always explicit and requires inferring connections.
  2. Dynamic Concepts: Concepts evolve and can be abstract, making them hard to map directly to financial data.
  3. Emerging Markets: Markets like China, historically less accessible, have less readily available structured information and analysis. Traditional methods often rely on direct evidence or extensive manual labeling, which is costly and may not capture the nuanced, multi-step reasoning needed to connect concepts to stocks.

Methods

The core of the paper's contribution lies in a reinforcement learning (RL) framework for concept expansion.

  • Representation: They leverage word embeddings (trained on large text corpora) to represent concepts, stocks, and documents, capturing semantic relatedness. Local embeddings (trained on data from specific sources) are used for more nuanced query expansion.
  • Concept Expansion as an MDP: The process of expanding a concept is framed as a Markov Decision Process (MDP).
  • States: Represent the current concept and its expanded list, along with candidate concepts and their supporting evidence.
  • Actions: The RL agent can choose to add a candidate concept, reject candidates, remove the last added concept, or stop the expansion process.
  • Reward: The reward function is based on the improvement in Mean Average Precision (MAP) of the stock recommendation list after an expansion step, encouraging actions that lead to better stock rankings.
  • Neural Agent: A deep Q-learning agent (using Bi-LSTMs) learns the optimal policy for selecting expansion concepts, guided by a small amount of manually labeled data. This allows for dynamic, multi-step reasoning, unlike fixed-rule expansion methods.

Datasets

The paper uses two Chinese stock market datasets:

  • Jinrongjie: Contains 206 concepts with an average of 25.4 manually suggested concept stocks.
  • Tonghuashun: Contains 900 concepts with an average of 15.6 manually suggested concept stocks.

These datasets were collected from financial websites and manually analyzed by financial professionals, confirming their credibility. The authors also crawled stock lists from the Shanghai and Shenzhen Stock Exchanges, totaling 3326 stocks.

Four public data sources were used for evidence mining:

  • S1: News: Crawled from Sina Finance News (2009-2017).
  • S2: Reports: Annual and quarterly company reports from Sina Finance.
  • S3: Wikipedia: Relevant Wikipedia pages for concepts and stocks.
  • S4: Search Engine: Open-domain information from Bing API.

Key Insights and Ideas

  • Reinforcement Learning for Abstract Reasoning: Framing concept expansion as an RL problem allows the model to learn a flexible, iterative strategy for discovering indirect relationships between concepts and stocks, going beyond simple keyword matching.
  • Leveraging Local Embeddings for Nuance: While global embeddings capture general semantic relatedness, using local embeddings (trained on specific data sources) for query expansion is crucial for suggesting more relevant, context-specific concepts.
  • MAP as a Reward Signal: Using Mean Average Precision (MAP) as the reward function effectively guides the RL agent towards actions that improve the overall quality of the stock recommendation list, balancing precision and recall.
  • Data Efficiency: The RL approach requires significantly less training data compared to traditional ranking baselines to achieve competitive performance, making it more practical for emerging markets.

Key Results

  • Outperformance: The proposed RL method (denoted as RL) significantly outperforms all baselines, including a strong ranking baseline and various semantic-based methods, on both the Jinrongjie and Tonghuashun datasets.
  • Accuracy Metrics: RL achieved the highest scores across Precision@5 (P@5), Precision@10 (P@10), Recall@30 (R@30), and MAP. For example, on the Tonghuashun dataset, RL achieved a MAP of 0.378, compared to the next best baseline (Rank) at 0.356.
  • Data Efficiency: The learning curves show that RL requires less training data than the Rank baseline to reach similar performance levels.
  • Efficiency: While unsupervised methods like Search and Semantics are faster, RL is more efficient than the Rank baseline, which needs to rank every stock.
  • Data Source Effectiveness: News and Reports (S1 and S2) were found to be more effective data sources for improving recommendation accuracies during query expansion.
  • Qualitative Examples: The RL method successfully extended concepts like "Sino" with related concepts like "China" and "State-owned enterprises," leading to more accurate stock recommendations. It also demonstrated the ability to identify stocks based on indirect connections, though it occasionally made mistakes due to rumors or ambiguity.

Limitations and Future Work

  • Mistakes due to Rumor/Ambiguity: The RL method can sometimes select incorrect stocks due to rumors or ambiguous news, as seen with "Tesla" recommendations.
  • Comparison to Rank: While RL is more data-efficient, the Rank baseline still performs well and is simpler to implement.
  • Future Work: The paper suggests that the RL approach is capable of expanding concepts with relevant concepts, which helps in finding more relevant stocks. This implies that further exploration of concept expansion strategies could lead to even better recommendations. The authors also highlight the potential of their method for investors to gain understandings of emerging markets.

Summary updated

Abstract

We investigate the task of mining relevant stocks given a topic of concern on emerging capital markets, for which there is lack of structural understanding. Deep learning is leveraged to mine evidences from large scale textual data, which contain valuable market information. In particular, distributed word similarities trained over large scale raw texts are taken as a basis of relevance measuring, and deep reinforcement learning is leveraged to learn a strategy of topic expansion, given a small amount of manually labeled data from financial analysts. Results on two Chinese stock market datasets show that our method outperforms a strong baseline using information retrieval techniques.
BibTeX
@inproceedings{liu2018mining,
  author = {Qi Liu and Yue Zhang},
  title = {{Mining Evidences for Concept Stock Recommendation.}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2018},
  doi = {10.18653/v1/n18-1191},
}
Figure