In this thesis we investigate the forecasting problem for large-scale settings: how can we efficiently and accurately generate forecasts when we need to generate many of them? Motivated by the increasing availability of large volumes of data and the ever increasing popularity of neural network models we investigate how we can improve the efficiency and accuracy of neural network models for point- and probabilistic forecasting in Chapter 2. We find that we can achieve better forecasting accuracy whilst reducing resource consumption – leading to reduced operational costs – by designing a neural network that requires an order of magnitude fewer parameters as compared to existing neural network probabilistic forecasting models. However, we also see that outside academia, there is a different class of models that is commonly used to solve the point- and probabilistic forecasting problem at a larger scale. Thus, in Chapter 3, we investigate a similar question, but for a different class of models: how can we improve the efficiency and accuracy of Gradient Boosting Machines (GBM) models for point- and probabilistic forecasting? We propose Probabilistic Gradient Boosting Machines (PGBM), a method to create probabilistic predictions with a single ensemble of decision trees in a computationally efficient manner. We empirically demonstrate the advantages of PGBM compared to existing comparable state-of-the-art methods and find that PGBM can produce probabilistic estimates without compromising on point performance, using a single model only, thereby greatly improving forecasting efficiency as compared to existing methods. In Chapter 4 we investigate the problem of hierarchical forecasting, which is the forecasting problem where time series need to adhere to a cross-sectional or temporal hierarchy, for example product groupings at a grocery store. We find that existing hierarchical forecasting techniques scale relatively poorly to large-scale problem settings and propose to learn a coherent forecast for millions of time series with a single bottom-level forecast model by using a loss function that directly optimizes the hierarchical structure. This reduces the computational cost of the prediction phase in the forecasting pipeline, as well as its deployment complexity, whilst maintaining forecasting accuracy. We demonstrate the benefit in an offline test at bol, and show forecast improvements of up to 10%. Finally, in Chapter 5, we study another forecasting problem often encountered in industry: session-based recommendation. We investigate state-of-the-art methods for session-based recommendation and find that the most simple method gives the most accurate results. We propose Vector-Multiplication-Indexed-Session kNN (VMIS-kNN), an adaptation of a state-of-the-art nearest neighbor approach to session-based recommendation, which leverages a prebuilt index to compute next-item recommendations with low latency in scenarios with hundreds of millions of clicks to search through. Based on this approach, we design and implement a scalable session-based recommender system Serenade, which is in production usage at bol.