Högskolan i Skövde

This paper introduces a methodology for incorporating risk preferences directly into forecasting model selection. The relative model information score, estimated from either a point-based information criterion or cross-validated errors, leverages the full distribution to map different risk propensities. We show that standard model selection in the literature is risk-agnostic. A risk-neutral stance is represented by the median of the relative model information score distribution, which characterises the plausibility of a model choice, while risk-averse and risk-tolerant choices correspond to its upper and lower quantiles. Our empirical evaluation demonstrates that risk-neutral and risk-averse selections consistently outperform the benchmark risk-agnostic choice in both point and quantile forecast accuracy. Moreover, we show that a risk-tolerant selection is beneficial during periods of extreme disruption. The proposed methodology provides a robust and flexible way to manage the forecast modelling risk, improving forecast accuracy and aligning forecasting modelling with stakeholders’ risk profiles.

PRACTITIONER SUMMARY: In this research we introduce a methodology for incorporating stakeholders’ risk preferences directly in the forecasting process. Given a degree of risk-aversion, or risk-tolerance, our methodology can guide the selection of the appropriate forecast that exhibits these risk characteristics in its errors. Risk-averse forecasts will minimise the probability of large errors, while risk-neutral forecasts will focus on minimising errors irrespective of their magnitude. The standard model selection statistics in the literature are typically risk-agnostic and do not enable the analyst to implement a preference. Using data from retailing, we provide evidence that both risk-neutral and risk-averse forecasts outperform the standard solutions both in terms of point and quantile accuracy (from 4% to 82% depending on the case), with longer forecast horizons benefiting from increased risk-aversion. Additionally, we demonstrate that risk-tolerant forecasts are useful when there are large disruptions, resulting in reduced forecast errors. Our methodology is robust with respect to the choice of the exact level of risk-preference, therefore reducing the need to precisely elucidate the risk profiles of stakeholders, which can be challenging in practice. The proposed approach can be operationalised with any statistical, machine learning, or judgmental forecasts, allowing direct incorporation of user preferences in model selection. Moreover, it does not require any changes in the forecasting models, simplifying its adoption in practice. Our work constitutes a first step in incorporating stakeholder risk preferences into the forecasting process, and therefore into the decisions supported by these forecasts.