Risk Evaluation in a ML-Approximated Portfolio Environment

Abstract: This thesis explores and evaluates the forecasting application of the machine learning method Gradient Boosting Decision Trees. This method is used to forecast the demand of the online grocery market with a 7-day time horizon. The thesis was conducted in collaboration with the online grocery company Mathem. The model is applied and evaluated on three different periods representing the spring, summer and fall. The main evaluation metric is the mean absolute percentage error (MAPE), and clear differences were found depending on the predictability of the period. Apart from the model and its application to demand forecasting, the related risk was investigated. This was done by studying the Value-at-Risk and Expected Shortfall associated with discrepancies between the forecasted and actual values over the three periods. The most important conclusion of the case study at Mathem is that overestimation in the forecast is more costly in terms of monetary value than underestimating. It is also found that this is highly dependent on the cost structure of the company's operation and could therefore vary between companies. Thus, the study has contributed to understanding the applications of machine learning models in forecasting processes as well as the risks related to over/underestimating the demand of the online grocery market.