INFRANET: Forecasting intermittent time series using DeepNet with parameterized conditional demand and size distribution

Real-life time series data deals with the problem of intermittency, where the demand appears sporadically in time, i.e., long runs of zero demand are observed before periods with nonzero demand. Modelling time series for Intermittent demand forecasting (IDF) holds a significant role in several industries, such as Retail, Automobile, and slow-moving parts, to enable effective inventory management. Existing methods forecast inter-demand time and size independently, predicting constant values over the entire interval. This static forecasting reduces overall point accuracy. To tackle this issue, we introduce INFRANET (Intermittent demand Forecasting using Autoregressive Network), a novel approach for generating accurate probabilistic forecasts by parameterizing conditional demand probability and size distributions. We have proposed a novel framework incorporating an auto-regressive network that jointly models the demad probability and demand size via a weighted loss function. Additionally, we explore three new decoding techniques: two probabilistic and one deterministic approach for accurate demand forecasting. Our model is specifically designed to address intermittent and lumpy demand forecasting as well as obsolescence scenarios. We show that the proposed model outperforms existing methods, demonstrated through extensive empirical evaluation on multiple datasets.