TiDE: MLP-Based Long-Horizon Time-Series Forecasting

Summary

The article explains TiDE, a multilayer perceptron architecture for long-term time-series forecasting. It combines a history of the target series with historical and future covariates, encodes them into a dense representation, and decodes forecasts with a temporal component that can incorporate covariates relevant to individual future time steps. A linear residual path preserves a linear forecasting component, while shared weights let the model process separate series channels.

Key ideas

• TiDE uses dense MLP blocks to encode past observations and covariates, then decode future values.
• A temporal decoder reconnects future covariates to predictions at each forecast step.
• A linear residual path keeps a linear model within the architecture.
• The article reports favorable benchmark comparisons and faster training and production inference than Transformer models, but does not provide enough detail here to assess the evaluation setup.
• The practical MQL5 implementation differs from the paper's design, so its reported results may not transfer directly.

Tags

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