Uppsatsen skrevs våren 2022 för fastighetsbolaget Atrium Ljungberg. Syftet var att optimera personalfördelningen av Atrium Ljungbergs drifttekniker. Eftersom mängden personal som behövs beror på antalet inkommande ärenden så användes historisk data av anmälda serviceärenden för att skapa prediktiva modeller som prognostiserade ärendeflödet. I studien jämfördes statistisk tidsseriemodellering SARIMA, dynamisk regression (linjär regression och SARIMA) samt maskininlärningsmetoden XGBoost. Fulltext finns som pfd.
In order to keep a high standard in real estates owned by Atrium Ljungberg, they give their tenants the possiblity to communicate when there is an issue with the property they are renting. By optimising the staff according to the volume of real estate matters, the processing time can be minimised and the company can avoid over- and understaffing. This could result in more satisfied tenants and less workload for the caretakers. With the help of statistical methods the number of real estates matters could be forecasted. The purpose of this thesis was to build statisitcal models that could forecast the number of real estate matters in different operational areas, for the next 14 days.
The data that were used for analysis contained observations between november 2015 and december 2021 for the nine different operational areas. Three different types of models were compared to find the most optimal one for each operational area. The three types that were used was ARIMA, Dynamic regression models and XGBoost. The models were compared with RMSE and the one with the lowest RMSE for most forecast horizons was chosen as the best one. After the nine different models were chosen they were evaluated with MAE which measures the mean absolute error. This measurement is appropriate to use when studying the average number of errors made by the model.
The outcome showed that there were small differences in RMSE between the models within the different operational areas, which made it somewhat difficult to find the most ideal one. The conclusion was that XGBoost preformed best in five of the nine areas, and the dynamic regression model including two types of dummyvariables performed best in two of the areas. In the two remaining areas the dynamic regression model with one type of dummyvariabel and SARIMA model performed best.