Commodity Price Forecasting via Time-Series Returns augmented with Dense News-Derived Features

Forecasting commodity prices from news is challenging because the information contained in textual sources is often partially reflected in market prices. This thesis investigates whether text embeddings extracted from news can improve the prediction of the direction of Brent crude oil futures returns when combined with historical price information. The work first examines CrudeBERT as a standalone predictor, analysing its behaviour across different prediction horizons, market conditions and input languages. The results show that much of its apparent predictive power is explained by class imbalance in the return distribution rather than by genuine semantic understanding, while also highlighting the importance of language normalization and correct temporal alignment between news and market data. A forecasting pipeline combining price-derived features with both sentiment-based and embedding-based textual representations is then developed and evaluated using multilayer perceptron (MLP) and long short-term memory (LSTM) models. The experiments distinguish between an ideal setting, where same-day news is available, and a realistic forecasting scenario that uses only previously available information. Embedding-based features achieve about 60% directional accuracy in the former setting, while the best deployable configuration reaches approximately 53% average accuracy with balanced performance across classes. The results indicate that sentence embeddings provide more useful predictive information than sentiment probabilities and that the best performance is obtained only by combining textual and price-derived features. Beyond the forecasting results, the thesis provides a systematic methodology for evaluating news-based commodity price prediction systems under realistic operating conditions.