Project Abstract

Financial timeseries forecasting has an immediately identifiable application within the field of investing: knowing when a stock is about to appreciate or fall in value can be leveraged as an investment strategy. In the domain of financial prediction tasks, machine learning occupies an important place as an area of research and a set of tools for modelling practitioners. The Transformer machine learning model is one such tool and has been approached as a basis for a variety of enhancements. The modelling of certain companies’ performance could be improved via changes to the architecture, as well as a data-centric approach to capture both price signals and the dynamics of the economic context the companies operate in. Towards this end, we identify the Consumer Cyclicals industry as particularly likely to be influenced by the trends driving the wider economy. The suggest the kind of effect at play might be economic cyclicality, and to capture such economic signals, we propose an improved financial prediction timeseries dataset and a new way of tokenising data for a Transformer model. The approach seeks to improve the traditional timestep-wise timeseries tokenisation, which could remove interesting information about how dataset variates relate to each other. To enhance the dataset beyond the usual stock price timeseries utilised in timeseries forecasting research, alongside technical indicators and temporal data, we add a suite of economic indicators to it. We also augment the Transformer model’s ability to capture long-term trends by transposing input sequences. We hypothesise this way of conceiving of long-term signals as tokens helps preserve information about non-linear dependencies between dataset variates. We conduct a data ablation study on the dataset and experiments comparing the proposed architecture to baseline models, including Transformer and traditional machine learning models. We show that longer prediction tasks benefit from the addition of economic indicator signals to the training data. We also demonstrate the input transposition proposed by us has a beneficial impact of the performance of the Transformer model.

Keywords: machine learning, financial forecasting, timeseries forecasting

 Conference Details

Session: Poster Session A at Poster Stand 76

Location: Sir Stanley Clarke Auditorium at Tuesday 7th 13:30 – 17:00

Markers: Gary Tam, Betsy Dayana Marcela Chaparro Rico

Course: BSc Software Engineering FI, 3rd Year

Future Plans: I have a job lined-up