Conference Agenda

Time series forecasting has gained significant interest in recent years, not lastly due to its practical applicability in the form of demand forecasting in the field of logistics and e-commerce. Current methods frequently suffer from poor performance when applied to time series that exhibit a high degree of sparsity. However, in practice time series representing demands of goods in a warehouse for instance are highly sparse, with frequent time steps associated with zero demand. In literature, this problem is commonly referred to as Intermittent Demand Forecasting (IDF). In this paper, we will focus on one of the problems that arise in the context of IDF, namely that traditional time series loss functions are not well suited for the intermittent demand forecasting problem. Within this work, we will focus on identifying and resolving a number of shortcomings of an existing loss that has been specifically designed for IDF. Namely, Stock-keeping-oriented Prediction Error Costs (SPEC). We identify and propose solutions to a range of issues in aforementioned loss definition. We evaluate our method empirically by training models with our enhanced loss function on openly accessible benchmark datasets.

In this paper, we present a novel perspective on data filtering and present an innovative

wavelet-based approach that leads to improved ML based time series forecasts. A separation

of financial time series components into short-, mid- and long-run components allows us to

study the relevance of these frequency components with respect to a statistical accuracy

assessment for daily return forecasts.

A simulation study and an analysis of daily market prices suggest that particular short- and midrun

information components cover the relevant information that is necessary for estimating

precise daily forecasts. Furthermore, the deconstruction of a time series into different

frequency components in combination with Machine Learning results in superior forecasting

performance in comparison to econometric benchmarks.

The cumulative distribution function (CDF) of lead time demand stands at the heart of any inventory control application. Historically, the inventory control literature has concentrated on deriving theoretical distributions for lead-time demand and forecasting demand. Recent shifts in the literature advocate for a direct estimation of lead-time demand, either through estimating empirical distributions by the means of bootstrapping or by forecasting it directly (Babai et al. 2022, Boylan and Babai 2022). This research seeks to empirically assess these emerging methodologies, particularly in the challenging landscape of spare parts where both demands and lead times are sporadically observed. Moreover, we probe into the potential of machine learning, motivated by Bertsimas and Kallus (2020), as a tool to enhance the estimation of the CDF of lead-time demand, especially in scenarios with limited observational data.

We explore nonparametric conditional density estimation (CDE) tools (Bertsimas and Kallus 2020, Dalmasso et al. 2020). We posit that this density-centric approach provides a more detailed representation of the uncertainties surrounding lead-time demand within the spare parts context. We compare this approach against recently proposed bootstrapping and direct forecasting methods using real-world data, addressing the challenges associated with intermittent lead time observations.

This study bridges the classical paradigms focused on theoretical lead-time demand distributions and the emergent emphasis on its direct empirical estimation. Furthermore, by employing machine learning techniques, we want to develop a novel approach for estimating the CDF of lead-time demand, addressing the challenges faced in a spare parts inventory context with limited observational data.