M5 competition uncertainty: Overdispersion, distributional forecasting, GAMLSS, and beyond

The M5 competition uncertainty track aims for probabilistic forecasting of sales thousands Walmart retail goods. We show that the data faces strong overdispersion and sporadic demand, especially zero demand. discuss resulting modeling issues concerning adequate such count processes. Unfortunately, majority popular prediction methods used in (e.g. lightgbm xgboost GBMs) fails to address characteristics due considered objective functions. distributional provides a suitable approach overcome those problems. GAMLSS framework allows flexible using low dimensional distributions. illustrate, how can be applied by location scale parameter various distributions, e.g. negative binomial distribution. Finally, we software packages their drawback, like R package gamlss with its extensions, (deep) libraries as TensorFlow Probability.