Redes neurais aplicada ao turismo em Foz do Iguaçu

Abstract

This undergraduate thesis investigated the application of deep learning architectures, specifcally Long Short-Term Memory (LSTM) and Causal Convolutional Neural Networks (Causal CNNs), to forecast the number of formal employment ties in Tourism Characteristic Activities (ACTs) in Foz do Iguaçu, Brazil, from 2019 to 2023. The dataset, derived from the Annual Social Information Report (RAIS) provided by the Paraná Institute for Economic and Social Development (IPARDES), presented substantial challenges, including the short length of the time series (72 monthly observations), the presence of a severe structural break caused by the COVID 19 pandemic, and a high noise-to-signal ratio. The experiments were conducted in two stages: a Grid Search for hyperparameter selection and a Robustness Test to assess model stability under different random initializations. Although the LSTM networks achieved the best isolated results (R2 = 0.9062), neither architecture demonstrated statistical stability. Average R2 values were negative and showed high variance across runs, indicating that the models memorized noise rather than learning structural patterns. The Naive Baseline analysis outperformed the neural networks, suggesting that the end-to-end approach was unsuitable for this forecasting task. The study concludes that instability is the central finding, resulting from the combination of limited data and exogenous shocks. Future research should consider classical seasonal models (SARIMA), statistical decompositions (STL), and ensemble strategies to enhance predictive robustness.