Abstract
Demand forecasting of hierarchical components is essential in manufacturing. However, its discussion in the machine-learning literature has been limited, and judgemental forecasts remain pervasive in the industry. Demand planners require easy-to-understand tools capable of delivering state-of-the-art results. This work presents an industry case of demand forecasting at one of the largest manufacturers of electronics in the world. It seeks to support practitioners with five contributions: (1) A benchmark of fourteen demand forecast methods applied to a relevant data set, (2) A data transformation technique yielding comparable results with state of the art, (3) An alternative to ARIMA based on matrix factorization, (4) A model selection technique based on topological data analysis for time series and (5) A novel data set. Organizations seeking to up-skill existing personnel and increase forecast accuracy will find value in this work.
Introduction
In the high-stakes world of electronics manufacturing, predicting demand is critical for managing global supply chains where over $12 trillion in inventory is constantly in flux. Despite the availability of advanced machine learning techniques, many demand planners still rely on "judgmental forecasts" or simple statistical models. This reliance often stems from the perceived "black box" nature of state-of-the-art models. However, the cost of inaccuracy is significant: forecast errors for new products average 44-53%, leading to substantial stockouts or overstocking costs.
This paper addresses the challenges of scale and complexity in predicting demand for millions of hierarchical components and the difficulty of effectively incorporating "anticipatory data"—pre-orders that signal future demand before it actually happens.
Proposed Framework
The authors propose a cohesive system consisting of three main pillars: data preprocessing via Diagonal Feeding, prediction using Matrix Factorization, and automated model selection using Topological Data Analysis (TDA).
Diagonal Feeding
A key contribution is "Diagonal Feeding," a feature engineering technique that reshapes data to account for the timing of information availability. By aligning "h-period ahead" pre-orders into a diagonal matrix structure, the model captures cross-correlations between early signals and final delivery volumes. This approach achieved results comparable to models using 360 hand-engineered features, simplifying the pipeline for non-technical domain experts.
Matrix Factorization
The researchers adapted Temporal Regularized Matrix Factorization (TRMF) for demand forecasting. By factorizing the demand matrix of $n$ items over $T$ time steps into lower-dimensional latent factors, the model captures global patterns across different components that traditional local models, such as ARIMA, often miss.
TDA for Model Selection
To handle the scale of millions of time series, the authors employed Topological Data Analysis (TDA), specifically the Mapper algorithm. This technique clusters time series based on their topological features, allowing for the automated assignment of the most successful model (e.g., Adaboost vs. Random Forest) to each cluster. TDA proved highly scalable, processing 80,000 time series in under 30 minutes on a standard laptop.
Experimental Results
The study utilized a real-world dataset from a major electronics manufacturer involving 2,562 items over 45 time periods. The authors benchmarked 14 different methods using the Symmetric Mean Absolute Percent Error (SMAPE).
As shown in the results table, Adaboost emerged as the top performer with a SMAPE of 0.17, followed by an Ensemble model at 0.18. Both significantly outperformed the industry baseline, ARIMAX (0.32). The TDA-based selection pipeline demonstrated high efficiency, with a model selection error of only 6%. It correctly identified that Adaboost and Bayesian Structural Time Series (BSTS) were the optimal choices for 71% of the items.
Conclusion
This work bridges the gap between complex machine learning research and practical industrial application. By introducing Diagonal Feeding and TDA-based model selection, the authors provide a scalable, accurate, and interpretable framework for demand forecasting. While challenges remain regarding data sparsity and long-term forecast decay in matrix factorization, these contributions offer a robust roadmap for manufacturers looking to modernize their forecasting capabilities.
Reference: Rivera-Castro, R., et al. "An Industry Case of Large-Scale Demand Forecasting of Hierarchical Components." arXiv:2009.03262v1 [stat.AP].
Summary updated