Fair Walk-Forward Forecasting of ICE Robusta Futures: Comparing Statistical, Machine-Learning, Deep-Learning, and Hybrid Models

Abstract

Distribution shifts can degrade nonlinear forecasting models trained on historical price levels, particularly relative to models that retain the most recently observed level. This study empirically compares 14 statistical, machine-learning, deep-learning, and Autoregressive Integrated Moving Average (ARIMA)–recurrent hybrid specifications for the daily ICE Futures Europe Robusta Coffee benchmark, rather than proposing a new forecasting algorithm. The synchronized dataset comprises 4,603 observations from 14 January 2008 to 28 January 2026. Specifications are selected using a chronological pure-training/validation split and evaluated once on a 300-observation blind test set under a common one-step-ahead walk-forward protocol. The blind-test price distribution shifts sharply from gross training (Kolmogorov–Smirnov statistic 0.955; standardized Wasserstein distance 4.274). Delta-Random Forest achieves the lowest mean absolute error (MAE) at 83.59 USD per metric ton, whereas Khashei–Bijari ARIMA–Long Short‐Term Memory (LSTM) yields the lowest root mean squared error at 111.99. However, Delta-Random Forest improves MAE over Naive by only 0.27 USD per metric ton, and no selected pairwise comparison remains significant after Holm adjustment of the Diebold–Mariano tests. Grouped permutation importance shows that permuting lagged first-difference history increases Delta-Random-Forest MAE by only 0.76, while Volume and the three macro-financial regressors contribute negligibly. These findings support a cautious interpretation: differencing improves nonlinear learners’ robustness to level-range shifts but does not establish economically or statistically meaningful predictive information beyond the most recently observed price.

Received: 11 July 2026 | Revised: 14 September 2026 | Accepted: 23 September 2026

Conflicts of Interest

The authors declare that they have no conflicts of interest to this work.

Data Availability Statement

The frozen processed analytical dataset, preprocessing and modeling code, environment specification, validation-search tables, saved blind-test prediction vectors, and revision-only analyses are available in the public reproducibility repository: https://github.com/hdanginneu/ice-robusta-walk-forward-forecasting. Price and Volume were obtained from Investing.com, London Coffee Futures Historical Data (accessed January 28, 2026). The frozen analytical sample spans January 14, 2008, to January 28, 2026. Use of the provider’s data remains subject to its terms.  The macro-financial series were retrieved through FRED, Federal Reserve Bank of St. Louis: Nominal Broad US Dollar Index (DTWEXBGS); EFFR; crude oil prices: WTI—Cushing, Oklahoma (DCOILWTICO) (all accessed March 1, 2026). These source records are publicly accessible.

Author Contribution Statement

Phung Duy Quang: Conceptualization, Writing – review & editing, Supervision, Project administration. Cao Ha Hai Dang: Methodology, Software, Formal analysis, Data curation, Writing – original draft, Writing – review & editing, Visualization. Nguyen Huu Thinh: Conceptualization, Methodology, Validation, Writing – review & editing, Supervision. Hoang Nam Quyen: Investigation, Writing – original draft. Hoang Tuan Minh: Investigation, Data curation, Writing – review & editing. Van Thi Thuan: Investigation, Resources, Data curation, Writing – review & editing. Bui Thi Tam: Validation, Writing – review & editing, Visualization. Nguyen Viet Ha: Validation, Writing – review & editing. Tran Quang Thanh: Software, Validation, Writing – review & editing.