Prophet/forecaster.py:1789: RuntimeWarning: overflow encountered in exp

[MattiaPeiretti]

Hello, I am experiencing this issue while running prophet and scaling data. It also makes prophet significantly slower when data is scaled between floor and cap. About 10x slower.

prophet/forecaster.py:1789: RuntimeWarning: overflow encountered in exp
  sample_trends = cap / (1 + np.exp(-k_t * (t_time - m_t)))

Together with this I also get the following:

Optimization terminated abnormally. Falling back to Newton.

If I run prophet without scaling the data, I do not get these warnings.

Scaling data:

model_data['y_scaled'] = (model_data['y'] - model_data['floor']) / (model_data['cap'] - model_data['floor'])
model_data_scaled = model_data[['ds', 'y_scaled']].rename(columns={'y_scaled': 'y'})
model_data_scaled['cap'] = model_data['y'].max() * 2
model_data_scaled['floor'] = 0

model.fit(model_data_scaled)

What data is this about?

We're forecasting bank balances per day. Horizon is 90 days.

Prophet settings:
ModelType is logistic

The rest of the settings are tried in different combinations, but even with everything left to default, the problem still manifests itself.

Thanks in advance!

[imad24]

Please check the "Minmax scaling" section of this page: https://facebook.github.io/prophet/docs/additional_topics.html#:~:text=minmax%20scaling%20(new%20in%201.1.,-5