Our Client:

A private hedge fund

Situation:

Foreign exchange trading is a fast-paced environment in which traders require quick and accurate predictions of price movements to make informed decisions. Our client faced the challenge of predicting foreign exchange values using multiple data sources, including low-level order book data recorded at irregular time intervals, social media or news sentiment. Traditional time series models struggled to capture the complexities of these data types, making accurate predictions a daunting task.

The goal was to develop a state-of-the-art model capable of multi-horizon forecasting with high confidence, which could provide error bounds on the prediction to assist with algorithmic trading decisions. The model would incorporate a range of data sources to improve prediction accuracy and capture nuances of the foreign exchange market.

Task:

The main tasks involved the development of a machine learning model capable of predicting foreign exchange values using multiple data sources.

To achieve this, DSL utilized a state-of-the-art Temporal Fusion Transformer-based architecture with an extension capable of handling irregular time intervals and multiple data modalities. We designed a custom preprocessing pipeline to transform raw data into a format suitable for the model and incorporated additional external data sources.

The model was trained using historical data and optimized for multi-horizon forecasting and prediction error estimation. Finally, DSL backtested and integrated the model into the client's algorithmic trading system,providing predictions in real-time.

Result:

The model outperformed the previous approach by a significant margin in backtesting, demonstrating the effectiveness of the new architecture and multi-modal approach. The model was subsequently integrated into the client's trading system, where it successfully predicted foreign exchange values and aided investment decisions with good results.

The ability to incorporate multiple data sources, including social media and news sentiment, provided additional insight into the foreign exchange market and improved the model's overall performance, allowing the client to make confident trading decisions. The successful implementation of the model demonstrated the potential of state-of-the-art AI techniques in financial forecasting and trading applications.