Issue¶

Program stalls¶

https://github.com/facebook/prophet/issues/842

Too few data points lead to numeric singularity??? then use newton not L-BFGS. Other reasons???

Trend issue¶

The trend can be completely wrong for long term forecast: https://towardsdatascience.com/fixing-prophets-forecasting-issue-b473afe2cc70

Optimization terminated abnormally. Falling back to Newton.¶

https://github.com/facebook/prophet/issues/1678

by disabling the Newton fallback, will will see all the errors from stanpy

m = Prophet()
m.stan_backend.set_options(newton_fallback=False)
m.fit(df)

except RuntimeError as e: 

Line search failed to achieve a sufficient decrease, no more progress can be made¶

https://discourse.mc-stan.org/t/line-search-failed-error/10163

Very small alpha (related to step size in line search) supports
the assumption that the search direction is not pointing downhill.
You can also test running with smaller ODE solver tolerances.

lbfgs (CmdStan User's Guide: https://github.com/stan-dev/stan-dev.github.io/blob/master/users/documentation/index.md)

init_alpha = 0.001 (Default) The alpha value is for step size used
tol_obj = 1e-12 (Default)
tol_rel_obj = 10000 (Default)
tol_grad = 1e-08 (Default)
tol_rel_grad = 10000000 (Default)
tol_param = 1e-08 (Default)
history_size = 5 (Default)