Exploring Deep Double Descent in GBP/USD Forecasting

Summary

The document introduces deep double descent, a pattern in which test error can initially fall, rise near a model’s ability to fit the training data, and then fall again as training continues. It contrasts this with the usual overfitting picture and examines training duration as a variable, using GBP/USD daily data. The example searches combinations of return period and forecast horizon, evaluates linear regression with time-series cross-validation, then applies a neural-network regressor and describes a trading implementation.

The article argues that early stopping can prevent a model from reaching a later decline in test error, but presents the phenomenon as conditional and without a widely accepted explanation. Its trading account is described as volatile, and the conclusion acknowledges that the model was exposed to the data it was trained on. The author recommends a more robust holdout period, keeping later years unseen during training, so the reported behavior should not be taken as proof of out-of-sample trading performance.

Key ideas

• Deep double descent describes a possible second decline in test error after error rises near the interpolation threshold.
• The example studies training duration and forecast settings using daily GBP/USD data.
• Time-series cross-validation is used to compare return periods and forecast horizons.
• The article warns that early stopping may miss a later improvement in some tasks.
• The trading results are limited by volatile performance and insufficiently isolated holdout data.

Tags

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