Claim Missing Document
Check
Articles

Found 1 Documents
Search

Overcoming Tree Extrapolation Limits in Gold Price Forecasting: A Stationary Return-Reconstructed XGBoost Approach Imam Saputra; Zulfikar; Juni Ismail; Alfry Aristo Jansen Sinlae; Yanto Saputra; Raja Anan Nasution; Elsy Rahajeng; Mesran
Journal of Informatics, Electrical and Electronics Engineering Vol. 6 No. 1 (2026): September 2026
Publisher : Forum Kerjasama Pendidikan Tinggi (FKPT)

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.47065/jieee.v6i1.2850

Abstract

This study addresses the fundamental tree extrapolation limit inherent in gradient-boosted decision tree ensembles when forecasting non-stationary financial time series during historical market rallies. Standard raw-price regression models fail to project unseen price regimes because decision trees assign static scalar predictions to terminal leaf nodes, capping predictions at historical maximums. To overcome this spatial limitation, we propose a Stationary Return-Reconstructed XGBoost framework that shifts the machine learning target from non-stationary spot prices to stationary daily relative returns before deterministically re-integrating predicted returns into nominal price levels. Evaluated on an out-of-sample testing set spanning 2022 to 2024, the return-reconstructed framework completely eliminates flatline prediction plateaus during record-setting gold price rallies exceeding 2,800 USD per troy ounce. The proposed model achieves an R-squared score of 0.9905, a Root Mean Squared Error of 28.17 USD per troy ounce, and a Mean Absolute Percentage Error of 1.12 percent, outperforming the raw-price baseline model by 92.75 percent in average error reduction. Furthermore, game-theoretic SHAP analysis validates that daily return forecasts are driven by macroeconomically sound transmission channels, predominantly led by the US Dollar Index and monthly historical price memory