Reading financial news three questions at a time
Financial news, read sentence by sentence: which company is being talked about, when the event happens, and how it feels. Three separate questions.
Off-the-shelf financial sentiment tells you a headline is negative. It rarely tells you negative about whom, or whether the thing described has already happened. A single article can be optimistic about one company and grim about another, and can mix an event from last year with a forecast for next quarter. Treating that as one score throws away most of the signal.
Master's thesis, sole author, supervised by Dr Sandy Ingram. Everything from the dataset generation tooling to the six services and the interface.
No public dataset existed for sentence-level, finance-specific temporality, so the first deliverable is a tool that builds one.
Before any training, every sample is embedded by three independent encoders and the classes are tested for statistical separation.
An article is split into sentences and then into clause-sized fragments, and each fragment is bound to the companies it actually mentions.
Each fragment goes to two services in parallel: when is this happening, and what does it feel like.
Results are counted per company and per time window, and when the top two emotions are close the system reports the named pair they form.
Segmentation, entity binding, temporality, sentiment, the decision maker and the dataset CLI. Separated because their dependencies genuinely conflict
Rule-based date normalisation; needs Java, which is one reason it lives in its own container
One model for temporality, one for emotion, each trained on the generated corpus
Sentence splitting, part-of-speech tagging, and the custom entity model that binds fragments to tickers
The single entry point the interface talks to; aggregates and composes the dyads
Submit an article, read the analysis per company and per time bucket
instead of · Positive / negative / neutral, or predicting all the compound emotions directly
Ternary sentiment can't distinguish fear from anger, which matter differently to a market. Predicting every compound emotion directly means many more classes and worse performance on each. Training on the eight primaries and composing the pairs afterwards keeps the class count low and makes the composition rule defensible: it comes from Plutchik, not from me.
instead of · One model for all temporal expressions
A fixed set of weights cannot do calendar arithmetic against a moving today, since 2023 is recent in 2024 and ancient in 2030. So explicit dates go to a rule-based tagger, and the model only handles implicit language. The two halves don't overlap by construction: the generated training data was explicitly instructed never to mention specific dates.
instead of · Letting the model own what 'recent' means
A day trader's present is hours; a pension fund's is a quarter. If those thresholds live in the weights, serving both audiences means relabelling and retraining. In a config file it is a one-line edit. This is the decision I'd defend hardest.
instead of · A one-off generation script
Every existing corpus was the wrong granularity, the wrong domain or not public. Making the generator generic, driven by a TOML spec, meant the same tool produced both datasets and can produce the next one. The validation step was built into it rather than bolted on.
Off-the-shelf financial NLP tells you a headline is “negative”. It rarely tells you negative about whom, or whether the thing described already happened. SAFE splits an article into sub-sentence segments, detects the tickers mentioned, and then classifies each segment twice: on a five-class temporal axis (distant past to distant future) and on the eight primary emotions of Plutchik’s wheel. The results are aggregated per company per time bucket, and when two emotions come out close together the system reports the named dyad they form instead of picking a winner. Anticipation plus joy is optimism.
Temporality is deliberately hybrid: Stanford’s SUTime handles explicit dates with rule-based precision, and a fine-tuned BERT takes over only when no date phrase exists. The class boundaries are configuration, not training data, so what counts as “present” can be changed without touching the model.
No dataset existed for either task, so a second deliverable is a CLI that generates and augments labelled data with GPT-4 from a TOML spec and a handful of human-written seed sentences. Before any training, every generated set is validated: each sample is embedded with three independent encoders and pairwise t-tests check that the classes are actually separable. That pass is what caught fear and sadness overlapping under one encoder, a weakness worth knowing before, not after, training.
The honest limitation, stated in the thesis itself: the reported accuracies are validation figures on GPT-4-generated data, with no independent test set. They measure fit to a synthetic distribution, not performance on live market news. And the NER stage only covers NASDAQ listings.