UTFaculteitenEEMCSDisciplines & departementenDMBAssignmentsFinished AssignmentsFinished Master Assignments temp[M] Predicting Demand of Replacement Cars for Breakdown Cases using Machine Learning Techniques

[M] Predicting Demand of Replacement Cars for Breakdown Cases using Machine Learning Techniques

Master Assignment

predicting demand of replacement cars for breakdown cases using machine learning techniques

Type: Master M-BIT 

Location: University of Twente

Period: Jan, 2018 - Aug, 2019

Student: Salamah, S.Y. (Siti Yaumi, Student M-BIT)

Date Final project: August 28, 2019

Thesis (Restricted until August 31, 2024)

Supervisors:


Abstract:

Roadside assistance is an emergency service provided to assist people with vehicle breakdown incidents on the breakdown location. Rental cars are usually provided as replacement vehicles for vehicles that cannot be repaired on the spot. Predicting the demand of replacement cars for breakdown cases are essential for the rental car company providing this service. In this research, we aim to investigate to what extent the demand of replacement cars can be predicted. Domain analysis based on literature study and interview with domain experts were conducted to generate a list of potential predictors. Using real world data of replacement car orders in the Netherlands and external data such as weather and calendar data, we compared several machine learning and classical time series models to predict daily demand of replacement cars. Various aggregation levels for spatial level and product type were investigated. The result shows that the best performing model is an XGBoost model trained on a shuffled training and test set, with a 9.49% mean average percentage error. Moreover, we found that prediction performance gradually decreases as the prediction level goes deeper. In addition, we proposed to address the outlier demand by identifying outliers, predicting them separately, and classify a future observation. Empirical comparison of three different approaches was also carried out to produce prediction interval as a means to estimate uncertainty.