This paper focused on the task of sentiment analysis within the financial domain, aiming to classify text into positive, negative, or neutral sentiments. Employing an ELECTRA-small model initially pre-trained as a general sentiment classifier, a baseline model was trained on financial sentiment data, achieving an accuracy of 0.8547 on the Financial Phrasebank dataset. Misclassification of data between positive and neutral sentiment classes was the most pronounced cause of error. While attempts to augment the model’s financial vocabulary using the Fin RAD dataset led to decreased model accuracy, the introduction of adversarial attacks proved to be successful in improving the performance of the baseline model. Particularly, the model trained on data augmented with Text Fooler-generated adversarial examples exhibited a 4.68% increase in accuracy to 0.9015. This approach also reduced misclassifications between positive-neutral classes, thus mitigating the major challenge observed in the baseline model. This result is significant considering how well the model generalized to a challenging problem on a dataset that it never encountered before and sets this result apart from other contemporary work in literature which uses a subset of Financial Phrase bank as training data for fine-tuning.