The best approach is to have some kind of physical model of what you think is going on that can serve as a basis. For example, is it a downward trend with noisy measurements or can the trend actually plateau? Either way, you can model this with prophet’s additive model (which is very simple and would be easy to translate to PyMC if someone hasn’t done that already). You can come up with multiple models and with this much data compare them.
One way to do this is to have an underlying time series with a negative trend and then noisy measurements around the trend. The trend can be modeled parametrically or non-parametrically. You can model the underlying trend for all dates with sparse measurements, or in some cases, for example with normal noise, you can analytically figure out the change over multiple days (e.g., if the process is a random walk with standard normal change each day, the effect of two days is the sum of two standard normals, the variance of which is the sum of the normals and the location of which is the sum of the locations (often assumed to be zero).