5. Conclusion, Limitations, and Future Work
The development of an accurate classifier of investor sentiment is required to support decision making in financial markets. Using data from StockTwits, it is shown that MaxEnt and NB outperform SVM despite their simple classification foundation with a strong independence assumption of the features. Moreover, bi-grams and tri-grams robustly boost the classification performance of investor sentiment, capturing long-range dependencies to some extent in the tweets. Although negation is one of the key grammatical rules that inverts the meaning and polarity of a sentence in multiple ways, the implemented negation tagging mechanism does not lead to significant improvement in the performance of classifiers (see Figure 3). Besides, the domain-specific lexicon does not illustrate expected performance in our dataset, confirming its limited ability to capture complex linguistic structures and entities. As discussed in Section 4.1.2, this study reveals that emojis carry very strong discriminative power in the finance context in spite of their domain-specific pattern of usage. Thus, the existence of emojis in the financial texts has contributed substantially to classification performance. In general, deep neural networks significantly outperform traditional methods, depending on the topology and word embeddings (see Figure 7). As we have discussed previously, LSTM and GRU unexpectedly perform better than CNN, although the StockTwits messages are quite short and therefore suitable for CNN to learn local features. LSTM demonstrates robust ability to capture long-term discriminative dependencies without any feature engineering. It is able to focus on the linguistic entities such as emojis, negation, and sarcasm to some degree.
