Time Series, Stocks and News: A Multi-Model Forecasting Approach for the Infotech Sector

This thesis explores the efficiency of different forecasting techniques in predicting financial data in the infotech industry by leveraging five different key assets: NVDIA, Apple, Broadcom and Microsoft stocks, as well as an ETF tracking the whole sector (IUIT.L). Trying to address the sector’s strong trend and fast shifts, the analysis includes classical time series models (such as ETS, ARIMA and GARCH), newer trend-seasonality models (such as Prophet) and more advanced machine learning approaches (such as XGBoost). Including a dataset containing over 62,000 news sourced from the New York Times, this analysis tries to address whether a multi source model can improve forecast accuracy. Methods include sentiment extraction to generate features that can be included in the forecasting process. A full analysis was dedicated to understanding the intricate relationship between sentiment scores and asset’s returns. Results highlight a strong growing trend characterizing the sector, stopping during the first period of 2025, while unpredictability increased during the same time window. To address drastic changes in direction, the changing volatility underlined during GARCH implementation, and short term movements, XGBoost was found to be the best model for the scope. When implementing Sentiments, a complex non-monotonic relationship emerged. By injecting different scenarios in the Sentiment data, context specific influences were highlighted. The inclusion of sentiment scores between the model’s features (even if theoretically promising) did not outperform the financial data alone. This outcome can be linked to the increased variability brought by the sentiment scoring algorithm and a data source for news that is too generic for the scope.