The study explores the predictive power of Donald Trump's tweets on stock price changes by employing document embeddings and a neural network model, highlighting the limitations and opportunities within Twitter-based financial predictions. The research addresses past incidents that questioned the reliability of Twitter sentiment in trading, such as the 2013 Associated Press hack, and aims to improve upon these models by leveraging contextually-rich sentence representations. The model uses doc2vec embeddings and a neural network to predict stock changes, focusing on the period from 2010 to 2019 and testing its effectiveness on various financial securities, particularly government-contracted companies and technology firms. The findings suggest that while the model is adept at predicting the direction of price changes, it struggles with the scale, and that government-contracted companies yield more consistent profits, likely due to their dependency on presidential actions. The study concludes by acknowledging the potential for more granular analysis with microscale stock data and suggests that future research could benefit from applying pre-trained word embeddings like BERT and GloVe to refine predictions further.