Publicação
Automatic forecasting of bike-sharing demand
| Resumo: | This thesis aims to compare the forecasting accuracy and computational efficiency ofthe time series forecasting models SARIMA, Theta and Prophet based on the data from2017-2019 of a bike-sharing company from San Francisco. The approach of the experimentwas to do a rolling-origin forecast for univariate data to guarantee a suitable comparisondespite the models differences. The results have shown that each model had each ownstrengths. SARIMA excelled at portraying seasonality and overall having the highestaverage forecasting accuracy, its amount of preprocessing steps and computational costsare significantly high, which makes it unsuitable for automatic forecasting. Theta has nopreprocessing steps and produces forecasts almost instantly due to its simplicity, but hasconsiderably worse forecasts in comparison to the other two models. Moreover, both modelsstruggle with outliers and the impact of external covariates, which makes it difficult to applyto bike-sharing demand forecasting. Prophet proves to have a decent balance betweencomputational efficiency and precision. Even though the model is highly inaccurate inthe beginning of the rolling forecast approach, it manages to improve significantly to acomparable accuracy of SARIMA over time and furthermore succeeds at dealing withoutliers and additional external factors. The findings show that there is a trade-off betweenforecasting accuracy and computational efficiency and the ideal balance and model dependson the needs of the business. Future work could include the automation of SARIMA’s neededpreprocessing steps, further testing on additional datasets or the inclusion of additionalvariables. |
|---|---|
| Autores principais: | Eid, Amir |
| Assunto: | Univariate time series Automatic forecasting Bike-sharing demand Séries cronológicas univariadas Previsão automática Procura de bicicletas partilhadas |
| Ano: | 2026 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Católica Portuguesa |
| Idioma: | inglês |
| Origem: | Veritati - Repositório Institucional da Universidade Católica Portuguesa |
| Resumo: | This thesis aims to compare the forecasting accuracy and computational efficiency ofthe time series forecasting models SARIMA, Theta and Prophet based on the data from2017-2019 of a bike-sharing company from San Francisco. The approach of the experimentwas to do a rolling-origin forecast for univariate data to guarantee a suitable comparisondespite the models differences. The results have shown that each model had each ownstrengths. SARIMA excelled at portraying seasonality and overall having the highestaverage forecasting accuracy, its amount of preprocessing steps and computational costsare significantly high, which makes it unsuitable for automatic forecasting. Theta has nopreprocessing steps and produces forecasts almost instantly due to its simplicity, but hasconsiderably worse forecasts in comparison to the other two models. Moreover, both modelsstruggle with outliers and the impact of external covariates, which makes it difficult to applyto bike-sharing demand forecasting. Prophet proves to have a decent balance betweencomputational efficiency and precision. Even though the model is highly inaccurate inthe beginning of the rolling forecast approach, it manages to improve significantly to acomparable accuracy of SARIMA over time and furthermore succeeds at dealing withoutliers and additional external factors. The findings show that there is a trade-off betweenforecasting accuracy and computational efficiency and the ideal balance and model dependson the needs of the business. Future work could include the automation of SARIMA’s neededpreprocessing steps, further testing on additional datasets or the inclusion of additionalvariables. |
|---|