SK C&C · Jul 2019 – Mar 2020
Market Caster: News-based Market Prediction
01Background & Goals
- An AI and big-data platform that forecasts market indicators such as global indices and bonds over several horizons.
- Structured data was analyzed as time series, while financial events were extracted from unstructured news for prediction.
02Contributions
Event extraction and embedding
- Extracted the core components (actor, action, object) of news sentences
- Embedded event tuples with a neural tensor network and analyzed sentence similarity
Prediction models
- Predicted market events with a hierarchical event- and time-level attention model
- Built news sentiment modules (BERT, Naive Bayes) and ensembled structured- and unstructured-data models to predict market moves
03Tech Stack
- Framework / Platform
- PyTorch, scikit-learn
- Methodology
- Hierarchical attention, Neural tensor network, Sentiment analysis