Training Recurrent Autoencoders to Preserve Time-Series Memory

Summary

This article explains why recurrent networks may suit market time series: they carry a hidden state forward, allowing information from earlier observations to influence later processing without repeatedly supplying a large fixed input window. It describes an unsupervised recurrent autoencoder approach, with recurrent LSTM blocks in the encoder and fully connected layers in the decoder. The training setup is designed to make the model use information from previous iterations instead of reconstructing each input independently and discarding its memory.

The author implements the model in MQL5 with OpenCL and reports a test in which it reconstructs recent input data with a loss rate below 9%, while retaining information from at least 30 prior iterations. The article presents this as evidence that the proposed training approach is workable. These results concern reconstruction, however; they do not establish predictive power or trading profitability. The author also says that experiments forecasting the next price bar were not conducted, so the practical value for market decisions remains untested.

Key ideas

• A recurrent network can pass a compressed hidden state between time steps to retain information from earlier observations.
• A recurrent autoencoder must be trained so reconstruction depends on previous iterations rather than only the current input.
• The described design uses LSTM blocks in the encoder and fully connected layers in the decoder.
• The reported test measures reconstruction quality and retained history, not trading returns or forecast accuracy.

Tags

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