ARIMA?!

• The problem

• Is it a forecast problem?

• Weakly seasonality

• Multiseasonal time series

• Machine learning approach

Forecasting the number of visitors to Nettebad Osnabrück. Using

• Visitors to the pool from 2005-03-20
• Some variables about the pool such as events, classes, availability of certain facilities etc.
• Weather data

https://inclass.kaggle.com/c/swimming-pool-visitor-forecasting

Is it a forecast problem?

• It seems so, but looking at the lag plots ...

• It seems so, but looking at the lag plots ...

• It seems so, but looking at the lag plots ...

• It seems so, but looking at the lag plots ...

• And the autocorrelation plot

• Autocorrelation plot suggested weekly seasonality in the data
• The missing days is imputed and the time series is decomposed

• One could remove seasonality and do the prediction on the remainder, then add seasonality

• The best RMSE I could get with this approach was 330.15

• As there are two kinds of sesonality, one can use multiseasonal time series and TBATS

• Reached RMSE of 376

• School and bank holidays
• Weekday and month name to consider seasonality
• Weather data (temprature, wind, preception,...)
• New features
  • Monthly average temprature
  • Warmer than monthly average
  • Warmer than the previous day
  • Heat index
• Adjust prices by consumer price index (CPI)

• Train a random forest to get feature importance
• Use the most important feature (99% cumulative importance)
• Use gradient boosting (XGboost) for the final prediction

• Adjust christmas and new years manually to the previous year value

• The final rmse is 247.33 (269.89 without manuall adjusment)