Despite the vast advances in quantitative forecasting methods and their prevailing use in practice, expert judgment remains crucial in forecasting; either by specifying and supervising models, or by adjusting model forecasts using expert knowledge and contextual information. The latter is typically done outside any Forecasting Support System (FSS) since, unfortunately, many FSSs neglect judgmental interventions. This can complicate the job of analysts and makes the use of judgment challenging to track and evaluate. This thesis aims to explore how forecasters use contextual information to adjust statistical forecasts, narrowing our focus to demand forecasting.

First, we analyse a UK-retailer case study exploring its operations and forecasting process. Inspired by this case company, we describe laboratory experiments, simulating a demand planning process where both qualitative and quantitative information of unknown quality is presented to experimental participants. In the first study, we find that forecasters mainly focus on model-based anchors, yet they are also prone to react to overly optimistic qualitative statements. The results show the potential for participants to extract useful information even despite this task's complexity and information overload. One option for reducing the cognitive load is by structuring the contextual information. Hence our second study investigates whether the decomposition of qualitative information in the forecasting process with system support could be an effective tool for managing information overload and helping to weight it appropriately. Using a similar setup to the first experiment, we observe better forecasting performance when decomposition is employed. We also find it reduces the number and size of adjustments across all investigated treatments, which could counter some cognitive biases, such as the optimism bias.

While judgmental adjustments are the most common and well-documented form of expert interventions, in the case company we observe an alternative way of incorporating contextual information, which we call judgmental model tuning. Judgmental model tuning represents all changes that analysts make during the model building stage, such as adding features and changes to the model form. We analyse the use of judgmental model tuning, its effect on forecast accuracy, and contrast it with judgmental forecast adjustments. We find that model tuning is not as effective as we expect since demand planners tend to overuse it. Furthermore, we show that demand planners have the potential to identify and add important variables that are meant to capture special events into a forecasting model, but they do that inconsistently and too frequently. Nonetheless, our results are promising, since judgmental model tuning is cognitively easier, scalable, and can be adopted for multiple products simultaneously. The overall aim of this research is to understand how organisations utilise contextual information and algorithms and to design processes that better integrate these various sources of information to yield more accurate forecasts. This thesis contributes to these areas and identifies new future research questions that are important for effective use of human interventions in forecasts.