Foreign exchange prediction and trading using low-shot machine learning

Abstract

One Shot and Few/Low Shot Machine Learning are new novel techniques using less data in sequence learning for prediction analysis. This technique has been applied to image databases to further segment and create forecasting figures. In this paper, a financial dataset is converted and built into an image database of 5 feature classes. One shot and few shot learning models using prototypical networks and matching networks are tested on the built financial image database neural networks to forecast foreign exchange (Forex) rates, comparing the main trading currencies of Euro against US Dollar (EUR/USD). A comparison study has also been done using a built meta-learner Long Short Term Memory(LSTM) to forecast the same exchange rate. The paper also examines the tuning hyperparameters for both few shot learning and LSTM. Finally, LSTM results are compared against both One shot and Few Shot learning to test the effectiveness of the respective models, in which few shot model scored the highest accuracy.