Designing Rolling LSTM and TFT Experiments with Changepoint Features

Summary

This experiment runner configures repeated, rolling train-and-test evaluations for LSTM and Temporal Fusion Transformer models on a multi-asset Quandl dataset. It offers variants with different input sequence lengths and optional changepoint feature lookbacks, and builds project names from the architecture and experiment settings. Test intervals advance across years while sharing a configured training start year.

The script passes model parameters, a feature file, asset mappings, and minibatch-size choices into a function that runs each window. It also exposes settings for training and validation splits, time features, and diversified validation Sharpe evaluation. These are experiment configurations, not reported findings: the document includes no model results, benchmark comparison, data-quality discussion, or evidence that changepoint inputs improve performance. The code alone does not establish whether the evaluation avoids leakage or reflects live trading constraints.

Key ideas

• The runner compares LSTM and TFT architectures across rolling yearly test windows.
• Experiment variants toggle changepoint features with different lookback lengths.
• Sequence length, train-validation split, time features, and validation Sharpe settings are configurable.
• The script specifies an evaluation setup but provides no results or evidence of predictive improvement.
• The document does not establish whether the workflow addresses leakage or live execution constraints.

Tags

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