Improving sporadic demand forecasting using a modified k-nearest neighbor framework

Forecasting sporadic or intermittent demand presents significant challenges in supply chain management, primarily due to the frequent occurrence of zero demand values and the inherent difficulty in capturing underlying hidden patterns in sporadic dataset. Although various parametric and non-parametric methods are available for sporadic demand forecasting, they often fail to detect these hidden patterns, underscoring the need for artificial intelligence (AI) algorithms which are effective in identifying irregular patterns. Among AI algorithms, the k-Nearest Neighbor (k-NN) is particularly good at identifying patterns with limited data. In the past, k-NN based frameworks have been recognized for their ability to effectively forecast sporadic demand, particularly when it is anticipated that patterns will reoccur in subsequent non-zero demand values. However, this assumption does not always hold true in different real-world scenario. In sporadic time series data, zero values often comprise a significant proportion (>30%). To address this, this paper proposes a "zero-inclusive k-NN framework" that leverages both zero and non-zero demand data to identify patterns. The proposed framework offers two significant features: it enables industrial managers to utilize a large number of nearest neighbors and provides adaptability in demand-vector length. Numerical investigations with both synthetic and real datasets affirm the superior forecasting performance of the proposed k-NN method when compared to existing k-NN framework and conventional parametric benchmark methods. The implications of our findings extend to domains where sporadic or intermittent demand forecasting plays a vital role.