In this study, we employ a multidimensional approach to extract investor sentiment from social media data using a non-economic-related dictionary, namely, the NRC-Emotion Association Lexicon. Considering a vast number of short text messages from the financial microblogging platform StockTwits, we analyze eight different emotions contained in the words of each text message. Subsequently, we classify these posts as a bullish or bearish signal on the basis of their emotional profile using machine learning techniques to develop aggregated investor sentiment. We show that this classification outperforms comparable classifications based on non-economic or two-dimensional dictionaries in terms of accuracy and data efficiency. Consequently, we are able to predict intraday returns for the NASDAQ 100, underlining social media activity’s potential economic relevance for predicting financial market movements. As measuring investor sentiment from text data, especially from social media, has attracted widespread attention in prior studies, this paper contributes to the economic research in the field of text analysis by suggesting the advantages of a multidimensional analysis. By indicating three key factors for designing accurate field-specific dictionaries, our results emphasize the need for (more specific) emotion-based and economic-related dictionaries in future research.