Say you're building a sales prediction model to predict tomorrow's sales value, as well as the next 2 weeks of daily sales. The model is being trained using daily data for the previous 1.5 years, and it follows a strong weekly seasonality pattern.
Obviously once you are happy with model performance, would it be necessary to re-train the model everyday to capture data up to and including yesterday in order to get the most accurate prediction for tomorrow's sales value? Essentially, the model would be trained on a rolling 1.5 year data set to capture yesterday's sales value.
Or does it depend on the type of model being used? Whether you need to to completely re-train when new data is present
I can see why it would make sense to re-train your standard time series forecasting models (ARIMA etc.), but I can also understand more sophisticated models (Neural Networks etc.) may generalize well enough to not have to re-train everyday.
I am looking for an explanation of models where you would and wouldn't re-train when new time series data is present.