The stock market provides a rich well of information that can be split across modalities, making it an ideal candidate for multimodal evaluation. Multimodal data plays an increasingly important role in the development of machine learning and has shown to positively impact performance. But information can do more than exist across modes— it can exist across time. How should we attend to temporal data that consists of multiple information types? This work introduces (i) the MEANT model, a Multimodal Encoder for Antecedent information and (ii) a new dataset called TempStock, which consists of price, Tweets, and graphical data with over a million Tweets from all of the companies in the S&P 500 Index. We find that MEANT improves performance on existing baselines by over 15%, and that the textual information affects performance far more than the visual information on our time-dependent task from our ablation study. The code and dataset will be made available upon publication.

Recent work in psychology has shown that people who experience mental health challenges are more likely to express their thoughts, emotions, and feelings on social media than share it with a clinical professional. Distinguishing suicide-related content, such as suicide mentioned in a humorous context, from genuine expressions of suicidal ideation is essential to better understanding context and risk. In this paper, we give a first insight and analysis into the differences between emotion labels annotated by humans and labels predicted by three fine-tuned language models (LMs) for suicide-related content. We find that (i) there is little agreement between LMs and humans for emotion labels of suicide-related Tweets and (ii) individual LMs predict similar emotion labels for all suicide-related categories. Our findings lead us to question the credibility and usefulness of such methods in high-risk scenarios such as suicide ideation detection.