Thursday 20 March 2025
A new study has shed light on the effectiveness of various time series forecasting models in real-world applications, specifically in the hospitality industry. The researchers evaluated a dozen different approaches, ranging from traditional statistical methods to cutting-edge machine learning and deep learning techniques.
The goal was to predict hourly sales at thousands of restaurants across Germany over a 14-day horizon. To achieve this, they compiled a dataset that included weather data, calendar events, and time-of-day patterns. The models were then benchmarked against each other using a set of standard evaluation metrics.
The results showed that machine learning-based meta-models consistently outperformed the other approaches, leveraging Pandas Data Frames as input for training. These hybrid solutions combined the scalability of Spark with the flexibility of Pandas, making them well-suited for large-scale deployment.
Foundation models, such as Chronos and TimesFM, also demonstrated promising performance. These transformer-based models were able to capture trends, seasonality, and special patterns directly from the data without requiring additional regressors or extensive feature engineering. While they did not quite match the accuracy of the machine learning-based models, their zero-shot inference capability made them an attractive option for environments with abundant GPU resources.
The study also highlighted the importance of scalability in real-world applications. The hybrid Spark-Pandas solution emerged as a practical and effective choice for large-scale deployment, particularly in environments where advanced computational resources were limited.
In contrast, traditional statistical models like Prophet underperformed compared to the other approaches. Deep learning models, such as N-Beats, showed promise but required extensive computational resources and feature engineering, making them less efficient for this use case.
The findings of this study have significant implications for industries that rely heavily on time series forecasting, such as finance, energy, and healthcare. By understanding which models are most effective in different scenarios, organizations can make more informed decisions about which approaches to employ.
The authors’ approach also highlights the importance of collaboration between academia and industry. By working together, researchers and practitioners can develop solutions that are both theoretically sound and practically applicable.
Overall, this study provides valuable insights into the strengths and limitations of various time series forecasting models. As the field continues to evolve, it will be important for researchers and practitioners to stay up-to-date with the latest developments and best practices.
Cite this article: “Time Series Forecasting Models in Hospitality Industry: A Comparative Study”, The Science Archive, 2025.
Machine Learning, Time Series Forecasting, Hospitality Industry, Germany, Weather Data, Calendar Events, Machine Learning-Based Meta-Models, Scalability, Deep Learning Models, Traditional Statistical Models
