The resurgence of deep learning has seen its
successful application to a variety of fields and industries.
However, its use in the prediction of financial markets remains
challenging as the associated data is noisy and the domain is
filled with nuance. Despite this, recent advances to feature
extraction techniques and model architecture have produced
promising results. In this paper, we first propose a novel multitask
regression model architecture using CNN, LSTM and the
correlation matrix of eight FX pairs to predict daily FX returns.
The use of the correlation matrix allows us to exploit both
temporal and spatial relationships between currencies. We then
compare our results to other baseline CNN based financial
prediction models. We show that our best models achieve a
RMSE of 6.38x10-3 and a profit of 6,126 pips during backtesting.
The work is done as part of the capstone project CS8-2 2020 S2.

свернуть