I used tbats to fit a model to a set of 3 years of historic data for daily number of shipments moved by a trucking company. my data included double seasonality so I used tbats. However, tbats did not consider holidays and the point forecast for holidays of the test data were very off from the actual data. This makes total sense as I did not include holidays in my forecast. Is it reasonable that I use this forecasting model but for holidays, I basically use the average value of that holiday day over the historic data as the point forecast? I tried arima with fourier series to include seasonality and regressors to include holiday but the fitted model was much worse than the tbats one. It worked better for holidays but was not accurate enough for other days.