Touristcast

A group project where we aimed to predict hotel stays from 2018-2019 for 13 regions in France. The presentation highlights project methodology, models used, results, and accompanying visualizations

Transcript

1. Touristcast Forecasting Hotel Stays in France Aysun Akarsu Sekou Camara Jinny To Franck Tramond github.com/jinnyto/touristcast
2. Datasets Hotel stays by region (INSEE) Regional weather (Historique Météo) Number of days off per month incl. weekends and holidays (Python holidays library) Regional gross domestic product (INSEE)
3. SARIMAX • Stands for Seasonal AutoRegressive Integrated Moving Average Exogeneous • It’s an Regression Technique which relates the value of an observation at time x, at time t, with some error e : ➔ How many periods in the past do we need to predict a value with minimal error?
4. Vector Autoregressive Capture linear relationships between multiple time series Time Series Vectorial System Input 1 Input 2 …. Input n Time Series Vector System Output 1 = f1(Input1) +f2(Input2) + ... Output 2 = g1(Input1) +g2(Input2) + ... …. Output n = h1(Input1) +h2(Inputn-1) + ... Input Output Each variable has an equation explaining its evolution based on his own lagged values, the lagged values of the other variables and an error term.
5. Prophet (Facebook AI) Adjustable Automatic Forecasting Procedure implemented by Facebook in 2017 Decompose and determine the model of a time series into 3 main components Trend Seasonality Holidays Error Term
6. Time Series Cross Validation - Have to take data points in order - Use only past data to predict future data - Walk-forward cross validation Years of training data: ['2010', '2011', '2012', '2013', '2014'] Predicted year: 2015-01-01 VAR lag order: 3 RMSE test: 58.63655909359081 MAE test: 46.72588709543521 Years of training data: ['2010', '2011', '2012', '2013', '2014', '2015'] Predicted year: 2016-01-01 VAR lag order: 3 RMSE test: 138.26758250655323 MAE test: 100.68670837522876
7. Model Comparison VAR vs. SARIMAX vs. Prophet
8. Performance Metric: Seasonal MASE
9. MASE - Best VAR, SARIMAX, Prophet models Performance using 6 periods of training data to predict 2017 hotel stays
10. RMSE - Best VAR, SARIMAX, Prophet models Performance using 6 periods of training data to predict 2017 hotel stays
11. Max error - Best VAR, SARIMAX, Prophet models Performance using 6 periods of training data to predict 2017 hotel stays 50%
12. SARIMAX Chosen model for 9 regions Bourgogne-Franche-Comté Bretagne Corse Centre-Val de Loire Grand Est Île-de-France Nouvelle-Aquitaine Occitanie Provence-Alpes-Côte d'Azur
13. Prophet Chosen model for 4 regions Auvergne-Rhône-Alpes Normandie Pays de la Loire Hauts-de-France
14. Seasonal Naive Hard to beat for 3 regions Bourgogne-Franche-Comté Bretagne Corse
15. Heatmap demo 2015 - 2019 - http://touristcast.appspot.com/2015_2019_heatmapwithtime.html
16. Appendix - Past Hotel stays by region
17. Appendix - Past Hotel stays by region
18. Appendix - Past Hotel stays by region
19. Appendix - Past Hotel stays by region
20. Appendix - Past Hotel stays by region