VideoConviction: A Multimodal Benchmark for Human Conviction and Stock Market Recommendations

Paper: VideoConviction: A Multimodal Benchmark for Human Conviction and Stock Market Recommendations Conference: ACM SIGKDD 2025 Authors: Michael Galarnyk, Veer Kejriwal, Agam Shah, Yash Bhardwaj, Nicholas Watney Meyer, Anand Krishnan, Sudheer Chava

Dataset Summary

VideoConviction is a multimodal dataset curated from financial influencer (“finfluencer”) videos on YouTube. It provides fine-grained annotations of investment recommendations, human conviction levels, and associated metadata. The dataset is designed to benchmark models that can jointly understand language, vision, and social signals to make sense of financial advice given by influencers.

This dataset facilitates research in areas such as:

• Investment recommendation detection
• Human conviction estimation
• Finfluencer behavior modeling
• Multimodal understanding and reasoning

Dataset Structure

Each row in the dataset corresponds to a specific segment within a YouTube video where a stock market action is recommended or discussed. The following features are included:

Primary Annotation Fields

• id: Generated ID for each video (from Label Studio).
• derived_inner_id: Generated inner ID for each video (from Label Studio).
• video_id: YouTube video identifier.
• start: Start time of the annotated video segment.
• end: End time of the annotated video segment.
• action: Suggested investment action (e.g., Buy, Hold, Don’t Buy, Sell, Short Sell, Unclear).
• action_source: Source within the video from which the action was extracted (e.g., Title, Introduction, Conclusion, Selected Region).
• conviction_score: Annotated conviction level of the recommendation (1/3 – low, 2/3 – medium, 3/3 – high).
• ticker_name: Stock ticker symbol (e.g., AAPL).
• action_date: Date associated with the recommendation.
• price: Recommended price to buy the stock at.
• quantity: Recommended number of shares.
• is_rec_present: Boolean indicating if the video includes any stock recommendation (Yes/No).

Video Metadata

• video_title: Title of the YouTube video.
• original_inner_id: Original inner ID (for data validation).
• original_video_title: Video title from YouTube metadata.
• publishedAt: Publication date of the video.
• channelId: YouTube channel ID.
• channelTitle: Name of the YouTube channel.
• videoDescription: Full video description.
• tags: YouTube tags.
• defaultAudioLanguage: Default audio language.
• duration: Full duration of the video.
• isCaptionAvailable: Boolean indicating if captions are available.

Social Engagement

• viewCount: Number of views.
• likeCount: Number of likes.
• favoriteCount: Number of favorites.
• commentCount: Number of comments.
• comments: Actual user comments on the video.

Channel Metadata

• channelDescription: Description of the channel.
• channelViewCount: Total views across the channel.
• channelSubscriberCount: Subscriber count.
• videoCount: Total videos on the channel.
• channelCategory: Category/type of content on the channel.

Multimodal Content

• transcript: Full transcript of the video.
• youtube_video_url: Direct link to the YouTube video.
• segment_transcript: Transcript of the specific annotated segment.

Citation

If you use this dataset, please cite our paper:

@inproceedings{galarnyk2025videoconviction,
  author = {Michael Galarnyk and Veer Kejriwal and Agam Shah and Yash Bhardwaj and Nicholas Watney Meyer and Anand Krishnan and Sudheer Chava},
  title = {VideoConviction: A Multimodal Benchmark for Human Conviction and Stock Market Recommendations},
  booktitle = {Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD '25)},
  year = {2025},
  location = {Toronto, ON, Canada},
  pages = {12},
  publisher = {ACM},
  doi = {10.1145/3711896.3737417}
}

Contact

For questions or collaborations, please contact:

• Agam Shah: ashah482@gatech.edu
• Veer Kejriwal: vkejriwal6@gatech.edu
• Michael Galarnyk: mgalarnyk3@gatech.edu