Intermittent demand is often marked by many periods of zero demand and occasional periods of non-zero demand, becoming a challenge for demand planners.
In this article, I examine the performance of four machine learning models — LSTM/RNN, SARIMA, XGBoost, and Croston — to estimate intermittent demand.
To evaluate these models, I conducted 40 rounds of time series simulations using two patterns:
20 series based on a lognormal distribution (featuring several zeros and low demand values with a right tail of high dispersion) and
20 series following a zero-inflated Poisson distribution (a very large number of zeros combined with a Poisson distribution with a low lambda).
The results indicate that the analysis of the coefficient of variation (CoV) of the series is a determining factor for the choice of planning strategy (MTS / MTO).
The article “Machine Learning Algorithms for Intermittent Demand” can be accessed and downloaded for free from the Academy:
Link: https://www.academia.edu/115121202/Intermittent_demand_analysis