This study evaluates the performance of Artificial Intelligence models, specifically Long Short-Term Memory and Convolutional Neural Networks, by comparing them with traditional econometric approaches, namely exponential smoothing (ETS) and SARIMA, for forecasting monthly tourist nights spent in Italy, distinguishing between domestic and international flows.
The models and algorithms are trained and tested on univariate time series of monthly tourist nights spent. Their predictive accuracy is assessed using the Mean Absolute Percentage Error. The analysis considers the temporal structure of the data, characterized by regular patterns of trend and annual seasonality (12 months), particularly evident in international tourist nights spent.
Although AI algorithms achieve satisfactory predictive accuracy on the test data, their performance tends to be inferior to Error Trend Seasonality (ETS) models when applied to univariate time series with strong regularity in temporal patterns. The results highlight that forecasts can serve as a valuable tool for supporting public policies, especially in improving the planning of tourism services during periods of high seasonality.
Beyond predictive accuracy, the paper emphasizes the importance of readability and interpretability of model outputs, which are key aspects in selecting forecasting methods for applied decision-making contexts.
