Textual data often accompanying financial data is a valuable but historically frequently underused source of information. This thesis builds on the Guided Topic Modeling method of Dangl and Salbrechter (2023), which measures firm-specific risk exposures from financial news but relies on non-context-aware word vectors (Word2Vec). The aim of this work is to replace these word vectors with context aware sentence embeddings, using a pipeline that can be built, retrained and operated on local hardware. The resulting pipeline consists of four steps and runs entirely on a single NVIDIA GeForce RTX 5090 GPU. First, a corpus of more than 94M Thomson Reuters news items is reduced to 12.36M articles through filtering, deduplication and cleaning rules. A tokenizer and three encoder models (ModernBERT and ChronoBERT) are then trained on this corpus and turned into sentence embedders by contrastive finetuning on 61,000 synthetically generated triplets. The resulting embedding space is used by GT Msbert, an adaptation of the original projection algorithm to embedded phrases and sentences. The results show that the gain from domain adaptation comes mostly from contrastive finetuning and less from pretraining. The topic clusters are sensitive to the wording of the seed input, but the cluster centroid, which is the quantity that drives the index, remains relatively stable. The resulting exposure measure correlates with established external indices (the Geopolitical Risk Index and the St. Louis Fed Financial Stress Index) at the aggregate level, and transfers to the firm and industry level given the topic is actually discussed in connection with individual companies.
en
Additional information:
Arbeit an der Bibliothek noch nicht eingelangt - Daten nicht geprüft
