UCL summer industry placement with Prof Peter Mitic, Head of Operational Risk Methodology at Santander UK. Co-developed a dataset of 2,022 news articles extracted from Refinitiv Eikon and News API for 20 of the 30 DJIA companies. The articles are manually labelled as relevant, quasi-relevant, or irrelevant based on their relation to the target company and split into 70% for training, 10% for validation, and 20% for testing. Applied NLP techniques to independently developed four hyperparameter-optimised RNNs (LSTM, Bi-LSTM, GRU, Bi-GRU) and seven fine-tuned encoder-only PLMs (BERT, DistilBERT, RoBERTa, DistilRoBERTa, ALBERT, I-BERT, MobileBERT) to detect irrelevant news articles to ensure only relevant articles are used for sentiment analysis. The accuracy, precision, recall, and F1 scores of these deep learning models, along with their confusion matrices and AUC-ROC curves, are compared to three traditional machine learning models (Multinomial Naïve Bayes, Logistic Regression, SVM). They are also evaluated for computational efficiency, focussing on their ability to run on the GPU in resource-limited environments as well as their training times, prediction times, and model sizes. DistilRoBERTa achieves a strong balance between performance and efficiency. It also performs well across varying training data sizes, outperforming the best RNN and traditional machine learning model with just 20% of the training data compared to the accuracy achieved by these models using 100% of the training data. When irrelevant news articles identified by DistilRoBERTa are removed, companies with higher proportions of such content experience significant sentiment shifts as indicated by FinBERT, while companies with lower proportions show more stable sentiment predictions. DistilRoBERTa’s generalisability is also tested on news articles from a DJIA company not included in the training set and on social media X posts from three non-DJIA companies.