Updating a fitted model
Forecasting is unique among predictive models in that the value of the data is its recency and each passing moment creates a new set of valuable data to use. A common situation with a forecast model is the need to refit it as more data comes in. The city of Baltimore, for example, may use the crime model to predict how many crimes they might expect to happen tomorrow, so as to better place their officers in advance. Once tomorrow arrives, they can record the actual data, retrain their model, and predict for the next day.
Prophet is unable to handle online data, which means it cannot add a single new data observation and quickly update the model. Prophet must be trained offline—the new observation will be added to the existing data and the model will be completely retrained. But it doesn't have to be completely retrained from scratch and the following technique will save a lot of time when retraining.
Prophet is essentially an optimization problem...
