Simple Recurrent Neural Networks for Forex Time-Series Forecasting

Summary

This article introduces simple recurrent neural networks (RNNs) for forex time-series forecasting and explains how their hidden state carries information across sequential inputs. It outlines the recurrent computation and backpropagation through time, including the vanishing-gradient problem that limits a simple RNN’s ability to learn long-range dependencies. Long short-term memory and gated recurrent units are mentioned as architectures designed to address that limitation.

The practical workflow uses Python and Keras to build regression and classification models, prepare sequential data, train the network, examine feature importance, and export a model for use in a MetaTrader 5 Expert Advisor through ONNX. The article proposes comparing the RNN with a LightGBM model from an earlier installment on the same data, and provides supporting model and preprocessing files. However, the supplied excerpt does not show the full experiment or its forecast and trading results. It also notes that the selected data omit certain lagged and differentiated variables, so findings would depend on those data choices and the evaluation setup.

Key ideas

• A simple RNN processes ordered observations using a hidden state that carries information from earlier steps.
• Backpropagation through time can produce vanishing gradients, limiting learning across long sequences.
• The article demonstrates a Python workflow for sequential forex regression and classification models.
• An ONNX export connects the trained model to a MetaTrader 5 Expert Advisor.
• The described LightGBM comparison and trading evaluation depend on data preparation choices, and results are not included in the excerpt.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.