A Comprehensive Forecasting Framework based on Multi-Stage Hierarchical Forecasting Reconciliation and Adjustment

Zhengchao Yang, Mithun Ghosh, Anish Saha, Dong Xu, Konstantin Shmakov, Kuang-Chih Lee

Abstract

Large-scale business forecasting requires maintaining mathematical coherence across hierarchical levels while preserving local accuracy. This paper introduces Multi-Stage HiFoReAd, a comprehensive framework that combines Bayesian-optimized model ensembles with a novel four-stage reconciliation process. By incorporating Hierarchical Harmonic Alignment via Fast Fourier Transforms (FFT), the framework intelligently preserves seasonal signals in noisy data. Evaluations on large-scale retail and traffic datasets show that Multi-Stage HiFoReAd significantly outperforms state-of-the-art transformer models and traditional reconciliation methods in both accuracy and scalability.

Introduction

In industrial contexts like Walmart's advertising platform, demand must be forecasted across a multi-level hierarchy—from national aggregates down to specific product items. This presents two major challenges: coherence (ensuring child forecasts sum to parent forecasts) and the accuracy-noise trade-off (stable top-level data vs. sparse, intermittent bottom-level data). Traditional methods like Top-Down or Bottom-Up reconciliation often sacrifice accuracy at one end of the hierarchy. Modern optimal reconciliation techniques like MinTrace often struggle with industrial scale or fail to preserve unique seasonal patterns.

Methodology: Multi-Stage HiFoReAd

The proposed framework, Multi-Stage HiFoReAd, is a modular system designed for distributed environments. The process is divided into two primary phases: modeling and reconciliation.

Distributed Base Modeling

Using Apache Spark, the system generates base forecasts for every node in the hierarchy. It employs an ensemble of LightGBM (LGBM), MSTL+ETS, and Prophet. To optimize performance for each unique time series, the ensemble weights are tuned using Bayesian Optimization (BO).

Overview of the Multi-Stage HiFoReAd framework architecture.

Multi-Stage Reconciliation and Adjustment

The core innovation lies in the four-stage reconciliation process:

  1. Top-Down (TD) Reconciliation: Establishes an initial baseline of coherent forecasts.
  2. Hierarchical Harmonic Alignment (HHAFA): This stage uses FFT to extract seasonality components. By calculating Jaccard similarity between parent and child seasonality, the system intelligently blends TD forecasts with BO-ensemble forecasts to preserve local seasonal "shapes."
  3. MinTrace WLS: The top three, more stable levels are reconciled using Minimum Trace with Weighted Least Squares to ensure mathematical coherence.
  4. Stratified Scale-Weighted Synchronization (SSW-FS): The bottom, noisiest levels are adjusted via scaling to maintain their relative weights while ensuring they sum to the stabilized higher levels.
Detailed breakdown of the four-stage hierarchical reconciliation process.

Experimental Evaluation

The framework was evaluated on four diverse datasets: Walmart Ads Demand (~12.8k series), the M5 Competition dataset, Road Traffic (PEMS-SF), and Tourism data.

Key Results

  • Accuracy: Multi-Stage HiFoReAd improved Absolute Percentage Error (APE) by 3% to 40% compared to the base BO-ensemble. It outperformed advanced transformer models (e.g., FEDformer, TiDE) by up to 92.9% in certain scenarios.
  • Scalability: Implemented on Spark, the framework processed large datasets in approximately 0.5 hours, significantly faster than traditional optimal combination methods which exceeded 3 hours.
  • Coherence: The system achieved 100% mathematical coherence across all hierarchical levels.

Discussion and Insights

A critical insight from this work is that seasonality is a signal worth saving. Standard reconciliation often "smears" unique seasonal signatures by forcing them to match parent trends. By using FFT and Jaccard similarity, HiFoReAd protects these signatures. Furthermore, the multi-stage approach acknowledges that different levels of a hierarchy have different statistical properties, requiring tailored reconciliation strategies.

While highly effective, the framework's complexity introduces computational overhead compared to simple methods. Future work will explore integrating deep learning directly into the reconciliation stages and handling "non-tree" hierarchies where products belong to multiple categories.

Summary updated

BibTeX
@inproceedings{yang2024comprehensive,
  author = {Zhengchao Yang and Mithun Ghosh and Anish Saha and Dong Xu and Konstantin Shmakov and Kuang-Chih Lee},
  title = {{A Comprehensive Forecasting Framework based on Multi-Stage Hierarchical Forecasting Reconciliation and Adjustment}},
  booktitle = {IEEE Big Data},
  year = {2024},
  doi = {10.1109/bigdata62323.2024.10825167},
}
Figure