Abstract
As Fintech continuously disrupt the finance industry, big data from non-traditional sources have been increasingly used to extract investment signals with automated solutions. We leverage a large-scale user post dataset from a financial social media platform and to predict investor sentiment using the natural language processing techniques and machine learning models. The BERT model built and fine-tuned on texts from finance domain outperforms existing baseline models and provides 86% of accuracy to correctly predict investor sentiment. Our results show that automated investor sentiment classification needs to built with both complex machine learning models and with good quality fiancnial data.