Adapting N-BEATS for Forex Forecasting with Quantile Loss
Summary
This article explains an attempt to adapt N-BEATS, a residual time-series forecasting architecture, for a forex Expert Advisor in MetaTrader 5. The model decomposes a series through stacked blocks that successively explain historical input and contribute forecasts. The implementation discussion covers manual gradient storage and backpropagation, SiLU activations, Adam optimization, and quantile loss to represent forecast uncertainty. It also describes monitoring for concept drift and using uncertainty in risk controls.
The article’s central finding is cautionary: directly applying N-BEATS to noisy forex data did not deliver the expected trading improvement. The author attributes the gap to market noise, missing market-microstructure considerations, and computational cost relative to the gains. While benchmark results for the general architecture are discussed, the excerpt gives no detailed trading metrics to substantiate the EA’s performance. The engineering infrastructure may support further experiments, but the proposed system is not presented as ready for live deployment; adaptation and comprehensive validation remain necessary.
Key ideas
- N-BEATS uses residual blocks to explain historical observations while producing forecasts.
- Quantile loss represents multiple points of a forecast distribution rather than only a point estimate.
- The implementation builds gradient propagation and Adam optimization for neural training in MQL5.
- The article reports that a direct forex adaptation did not produce the expected trading results.
- Noise, microstructure omissions, and compute costs are identified as obstacles to practical gains.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.