Hierarchical Ensembling of Heterogeneous Time Series Models for Retail Inventory Decisions: A Local Empirical Study on Two Public Datasets
Keywords:
retail forecasting, time series foundation models, forecast combinations, hierarchical shrinkage, inventory simulation, probabilistic calibrationAbstract
Whether improvements in retail forecasting translate into inventory benefits must be tested under common temporal splits and continuously evolving inventory states. This study develops InventoryFusion, a heterogeneous forecast ensemble that combines four lightweight baselines with Tiny Time Mixers R2.1, Chronos-2, and TimesFM 2.5 through a common P10/P50/P90 interface. Ensemble weights are estimated by shrinking validation losses across global, demand-segment, and individual-series levels. The evaluation uses 120 series from each of FreshRetailNet-50K and UCI Online Retail, arranged into three disjoint sampling cohorts per dataset. A 14-day horizon, a 7-day rolling step, and separated weight-fitting and test periods yield 1,560 test forecast windows and 21,840 series-origin-horizon observations. Hierarchical ensembling achieves WAPE values of 36.66% and 89.02%, compared with 36.51% and 93.81% for equal weighting. On UCI, hierarchical WAPE is 5.115% lower than equal weighting but 3.800% higher than validation-selected Chronos-2. Its central 80% interval covers only 54.720% of observations, exposing lower-bound miscalibration at zero sales. Paired whole-series cluster bootstrap intervals based on 2,000 resamples are complemented by hierarchy ablations, delayed-label online updates, stockout-flag sensitivity analysis, and strictly separated residual calibration. A continuous lost-sales simulation uses a 7-day review interval and a 7-day lead time to compare cumulative-residual inventory targets with sums of daily P90 forecasts across three holding-to-shortage cost ratios. The findings show that ensemble utility depends on the data distribution, the granularity of weight estimation, and protection-period calibration. Adding models or improving point forecasts does not automatically reduce inventory cost. Version-pinned local execution, series-level outputs, and traceable figures support examination of both the benefits and the limitations of the ensemble.
References
[1] Fotios Petropoulos; Xun Wang; Stephen M. Disney. The inventory performance of forecasting methods: Evidence from the M3 competition data. International Journal of Forecasting, 35(1), 251–265, 2019. https://doi.org/10.1016/j.ijforecast.2018.01.004
[2] J. D. Croston. Forecasting and Stock Control for Intermittent Demands. Operational Research Quarterly, 23(3), 289–303, 1972. https://doi.org/10.1057/jors.1972.50
[3] Ruud H. Teunter; Aris A. Syntetos; M. Zied Babai. Intermittent demand: Linking forecasting to inventory obsolescence. European Journal of Operational Research, 214(3), 606–615, 2011. https://doi.org/10.1016/j.ejor.2011.05.018
[4] Aris A. Syntetos; John E. Boylan; J. D. Croston. On the categorization of demand patterns. Journal of the Operational Research Society, 56(5), 495–503, 2005. https://doi.org/10.1057/palgrave.jors.2601841
[5] Xiaoqian Wang; Rob J. Hyndman; Feng Li; et al. Forecast combinations: An over 50-year review. International Journal of Forecasting, 39(4), 1518–1547, 2023. https://doi.org/10.1016/j.ijforecast.2022.11.005
[6] Pablo Montero-Manso; George Athanasopoulos; Rob J. Hyndman; et al. FFORMA: Feature-based forecast model averaging. International Journal of Forecasting, 36(1), 86–92, 2020. https://doi.org/10.1016/j.ijforecast.2019.02.011
[7] Abdul Fatir Ansari; Lorenzo Stella; Caner Turkmen; et al. Chronos: Learning the Language of Time Series. arXiv:2403.07815, 2024. https://arxiv.org/abs/2403.07815
[8] Abdul Fatir Ansari; Oleksandr Shchur; Jaris Küken; et al. Chronos-2: From Univariate to Universal Forecasting. arXiv:2510.15821, 2025. https://arxiv.org/abs/2510.15821
[9] Vijay Ekambaram; Arindam Jati; Pankaj Dayama; et al. Tiny Time Mixers (TTMs): Fast Pre-trained Models for Enhanced Zero/Few-Shot Forecasting of Multivariate Time Series. Advances in Neural Information Processing Systems, 2024. https://arxiv.org/abs/2401.03955
[10] Abhimanyu Das; Weihao Kong; Rajat Sen; et al. A decoder-only foundation model for time-series forecasting. Proceedings of the 41st International Conference on Machine Learning, PMLR 235, 10148–10167, 2024. https://proceedings.mlr.press/v235/das24c.html
[11] Google Research. TimesFM 2.5: google/timesfm-2.5-200m-pytorch model card and release notes. Hugging Face model card and official software repository, 2025. https://huggingface.co/google/timesfm-2.5-200m-pytorch
[12] Wei Yang; Defu Cao; Yan Liu. Foundation Models for Demand Forecasting via Dual-Strategy Ensembling. arXiv:2507.22053, 2025. https://arxiv.org/abs/2507.22053
[13] Yangyang Wang; Jiawei Gu; Li Long; et al. FreshRetailNet-50K: A Stockout-Annotated Censored Demand Dataset for Latent Demand Recovery and Forecasting in Fresh Retail. arXiv:2505.16319; accompanying Dingdong-Inc dataset, 2025. https://arxiv.org/abs/2505.16319
[14] Daqing Chen. Online Retail. UCI Machine Learning Repository [Dataset], 2015. https://doi.org/10.24432/C5BW33
[15] Chen Xu; Yao Xie. Conformal prediction interval for dynamic time-series. Proceedings of the 38th International Conference on Machine Learning, PMLR 139, 11559–11569, 2021. https://proceedings.mlr.press/v139/xu21h.html
[16] Rob J. Hyndman; George Athanasopoulos. Forecasting: Principles and Practice. 3rd edition. Melbourne: OTexts. Section 5.10, Time series cross-validation, 2021. https://otexts.com/fpp3/tscv.html
[17] Hans R. Künsch. The Jackknife and the Bootstrap for General Stationary Observations. The Annals of Statistics, 17(3), 1217–1241, 1989. https://doi.org/10.1214/aos/1176347265
[18] Dimitris N. Politis; Joseph P. Romano. The Stationary Bootstrap. Journal of the American Statistical Association, 89(428), 1303–1313, 1994. https://doi.org/10.1080/01621459.1994.10476870
Downloads
Additional Files
Published
Issue
Section
License
Copyright (c) 2026 Jiexiao Zhang
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.
Authors retain copyright and grant Discovery permission to publish their work as the original journal of publication. Accepted articles are made freely available under the Creative Commons Attribution–NonCommercial 4.0 International license (CC BY-NC 4.0).
The license permits non-commercial sharing and adaptation with appropriate attribution, a license link and identification of changes. Reuse beyond the license requires permission unless otherwise permitted by law. Follow any separate license or credit line for third-party material. Read the open access, copyright and fees policy for details.