Stacked Models for NZDJPY Price Forecasting and Residual Correction

Summary

This article builds a two-model forecasting approach for NZDJPY: an initial model predicts a future close, and a second model estimates the initial model’s residuals to correct its forecast. The author describes collecting minute-level prices, exploring trend and seasonal structure, comparing candidate regressors, and selecting stochastic gradient descent for the first stage. Feature selection methods and residual diagnostics are used to examine the data and model fit.

The residual model is a deep neural network, tuned through random search followed by an L-BFGS-B optimization attempt, then exported with the first model for use in an Expert Advisor. The reported comparisons say the tuned residual models beat the default neural network on validation data. However, residual autocorrelation remains, suggesting the first model may be misspecified or missing useful inputs. The article presents a particular workflow, not evidence of durable profitability; its conclusion acknowledges the unresolved residual behavior and the limits of interpreting or relying on the forecasts.

Key ideas

• The stacking setup pairs a future-price regressor with a second model trained to predict its errors.
• The author selects stochastic gradient descent as the first-stage model after comparing candidate regressors.
• Residual autocorrelation is treated as evidence that the model may not have captured all structure or may lack relevant inputs.
• A deep neural network is tuned in two stages to estimate the first model’s residuals.
• Validation comparisons do not establish persistent trading profitability, and the residual diagnostics reveal unresolved fit concerns.

Tags

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