Building an LSTM Recurrent Network for Financial Time Series

Summary

The article introduces recurrent neural networks as a way to process price sequences while carrying information forward between time steps. It explains why a fixed input window can miss events outside the selected span, then presents an LSTM block with memory and hidden-state streams. Forget, input, output, and candidate-content gates determine what information is retained, added to memory, and exposed as the current output.

It describes implementing the block in MQL5 and training it with backpropagation through time, summing gradients across the repeated steps that share weights. The authors compare its test behavior with earlier network types and report lower root mean square error and an upward trend in target-hit accuracy during training. They also note that the recurrent model requires more computation and implementation effort. The reported tests concern the article’s fractal-prediction setup; the excerpt does not provide enough detail to judge robustness, generalization, or live trading performance.

Key ideas

• Recurrent networks pass a hidden state forward to carry information across a time sequence.
• LSTM uses gated operations to control forgetting, memory updates, and output generation.
• The implementation trains shared recurrent weights by accumulating gradients across time steps.
• The article reports improved test error and training accuracy relative to earlier network approaches.
• The reported results are limited to the article’s fractal-prediction experiments and do not establish live trading performance.

Tags

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