Católica Lisbon School of Business & Economics
Permanent URI for this community
Browse
Browsing Católica Lisbon School of Business & Economics by advisor "Afonso, Pedro"
Now showing 1 - 6 of 6
Results Per Page
Sort Options
- Forecasting bike-sharing demand in Seoul : a comprehensive analysisPublication . Salerno, Federico; Afonso, PedroBike sharing programs represent the future of mobility, contributing to the creation of a ”green” economy and a more sustainable future. Providing the city with a stable and accurate supply of bicycles is a major concern for the city of Seoul. The objective of this research is to provide useful insights on how effectively forecast bike sharing demand through automated processes. Side goals are related with the difference between models’ performances as well as with drawing causal effects. Examining the public rented bicycles in the city, time series forecasting is imple mented through different methods, exploring both parametric and non parametric models such as Seasonal ARIMAs with exogenous variables, Multiple Linear Re gression and Support Vector Regression. This study takes into consideration mostly weather-related features, consistently with previous literature. Rides are intuitively influenced by features like temperature as well as by time effects that occur in certain periods of the year. After inspecting the different relationship between response variable and features, models were fit and tested. Consistently with regression errors measured on test set, SVR can be considered the best model for the aim of this research.
- Forecasting flight prices with machine learning models : a comparative analysis between low and high-cost airlinesPublication . Daly, Sophia Maria; Afonso, PedroForecasting fight prices is a challenging task due to the complex nature of the pricing algorithms that airlines use. Apart from the fact that these algorithms are not public, they have to take into account many different variables that affect ticket prices. Since the airlines’ demand forecasting may not always hold true as a result of varying demand, prices need to be adjusted accordingly. This approach is called dynamic pricing. It is a technique of price discrimination based on temporal differences mainly, leading to the widely spread assumption that the time of booking is a crucial determinant of the ticket price. This analysis shows that apart from days to departure, especially fight distance and airline type infuence the price significantly. That is, longer fights as well as fights operated by full-service carriers, as opposed to low-cost carriers, are usually more expensive. This thesis uses a dataset including the fight fares and other fight-related characteristics of one-way fights in the US between April and October 2022, retrieved from the search engine Expedia.com. The data is used to train and compare the performance of several supervised learning models aiming to forecast fight prices. Each model is deployed three times, first with the entire dataset, and then once with data only from low-cost-carrier and only from full-service-carriers, respectively. The most accurate models for all three datasets are the random forests followed by k-nearest-neighbor. The results of this thesis suggest that a large part of the fight price can be predicted using fight-related details such as days to departure and fight duration, yet, it also shows that there remains a certain inexplicable variability that could be due to external factors that are not included in the present analysis.
- House price prediction : a comparative analysis of machine learning approaches to study Melbourne’s marketPublication . Nobile, Simona; Afonso, PedroThis thesis work investigates the application of machine learning (ML) techniques for predicting house prices, a crucial task with widespread implications. In this scope, this work presents a literature review on state-of-the-art approaches and a practical experiment using a dataset of house sales in Melbourne, Australia. The analysis focuses on identifying key features for price prediction and assessing the performance of various ML algorithms. In fact, examining feature importance over time, it is possible to understand the dynamic nature of house price prediction.
- Optimizing recycling using automated image-based classification : a machine learning approach for improved waste managementPublication . Böhmer, Lili Freia Annemarie; Afonso, PedroThis thesis covers image-based waste classification using machine learning and discusses its impact on sustainable waste management. To identify the optimal model, the prediction performance of a DenseNet, a state-of-the-art Convolutional Neural Network, and DAtNet model are examined and compared to each other. The DAtNet integrates attention layers on the DenseNet architecture, inspired by the transformer model, known for its success in large language models. Moreover, the impact of transfer learning and augmentation on the test accuracy is analyzed. The performance of these models is evaluated across multiple datasets to exam ine their generalization capabilities. The findings indicate that while the DAtNet surpasses the DenseNet model in accuracy with large datasets, it faces difficulties with smaller datasets and requires significantly more time to preprocess the images. In contrast, the DenseNet consistently performs well and processes images more efficiently. Therefore, a DenseNet model is recommended for waste management fa cilities due to its reliability and lower computational demands. However, the further investigation and improvement of attention layers is encouraged. Additionally, the development of more practical, representative datasets is essential for the effective implementation of machine learning models in real world waste management. The deployment of this work could support the achievement of Sustainable Development Goals and the realization of zero-waste cities.
- Predicting business cycles with linear and non-linear filtersPublication . Abrantes, Beatriz; Afonso, PedroBusiness cycles represent the short-run fluctuations in economies and have a non recurring periodic character that makes them difficult to forecast. This dissertation fo cuses on the cycle-trend decomposition techniques that are used to remove the long-run component and thus obtain the cyclical component of macroeconomic series. Statistical filters can be used for this purpose, and through them, this work aims to clarify and visualize the cycle-trend decomposition. The primary objective of this dissertation is to evaluate the performance of two types of filters, linear and non-linear. At the end, it is also expected that conclusions will be drawn about the tool used throughout this work, Power BI. After comparing the linear filter developed by Hodrick and Prescott (1997) with two non-linear filters, MR filter and median filter developed by Mosheiov and Raveh (1997) and Wen and Zeng (1999), respectively, the results obtained were favorable compared to the non-linear filter. The MR filter proved to be able to produce a more robust trend than the others and to identify economic periods in a natural way. The MED filter proved to be able to produce less volatile and noisy cyclical components than the others; this is due to its ability to capture sharp changes in the trend and suppress them in the cyclical component. This concluded that the nonlinear filters performed well against the linear filter under study. Power BI demonstrated throughout the work several capabilities that characterize it as a good Business Intelligence tool, however, with room for improvement.
- Short-term solar forecasting in Germany : using satellite imagery and a hybrid CNN-2-LSTM approachPublication . Voß, Leonhard; Afonso, PedroThis thesis presents a novel hybrid convolutional neural network to long short-term memory (CNN-2-LSTM) model for short-term solar forecasting using data from Southeast Germany. The model’s performance is compared against a long short-term memory (LSTM) model and a persistence forecast. Satellite-derived cloud masks from the CLAAS-3 dataset by METEOSAT are combined with minute-byminute global horizontal irradiance (GHI) data from the Fraunhofer Institute’s PV-Live dataset. The models predict solar output over 15 minutes, 3 hours, and 6 hours with various input sequence lengths, resulting in 37 models tested. A new performance metric is introduced, using balance energy prices to calculate the economic impact of the models, providing a practical perspective on financial implications. Results show that the CNN-2-LSTM model significantly outperforms benchmarks at the 15-minute horizon, demonstrating superior accuracy for very short-term predictions. However, its performance at the 3-hour horizon is comparable to the LSTM model, and it lags behind the LSTM model at the 6-hour horizon. These findings highlight the model’s effectiveness for very short-term forecasting while emphasizing the need for optimization for specific temporal scales. The research underscores the potential of hybrid deep learning approaches to enhance short-term solar forecasting accuracy. It offers a cost-effective alternative to more complex systems, valuable for solar energy businesses. The study encourages further exploration into optimizing deep learning models for different forecasting horizons, contributing to advancements in renewable energy management in line with sustainable development goals (SDG).