Facebook-Prophet: Overflow error when fitting - facebook-prophet

Question

I wanted to practice with prophet so I decided to download the "Yearly mean total sunspot number [1700 - now]" data from this place
http://www.sidc.be/silso/datafiles#total.
This is my code so far
import numpy as np
import matplotlib.pyplot as plt
from fbprophet import Prophet
from fbprophet.plot import plot_plotly
import plotly.offline as py
import datetime
py.init_notebook_mode()
plt.style.use('classic')
df = pd.read_csv('SN_y_tot_V2.0.csv',delimiter=';', names = ['ds', 'y','C3', 'C4', 'C5'])
df = df.drop(columns=['C3', 'C4', 'C5'])
df.plot(x="ds", style='-',figsize=(10,5))
plt.xlabel('year',fontsize=15);plt.ylabel('mean number of sunspots',fontsize=15)
plt.xticks(np.arange(1701.5, 2018.5,40))
plt.ylim(-2,300);plt.xlim(1700,2020)
plt.legend()df['ds'] = pd.to_datetime(df.ds, format='%Y')
df['ds'] = pd.to_datetime(df.ds, format='%Y')
m = Prophet(yearly_seasonality=True)
Everything looks good so far and df['ds'] is in date time format.
However when I execute
m.fit(df)
I get the following error
---------------------------------------------------------------------------
OverflowError Traceback (most recent call last)
<ipython-input-57-a8e399fdfab2> in <module>()
----> 1 m.fit(df)
/anaconda2/envs/mde/lib/python3.7/site-packages/fbprophet/forecaster.py in fit(self, df, **kwargs)
1055 self.history_dates = pd.to_datetime(df['ds']).sort_values()
1056
-> 1057 history = self.setup_dataframe(history, initialize_scales=True)
1058 self.history = history
1059 self.set_auto_seasonalities()
/anaconda2/envs/mde/lib/python3.7/site-packages/fbprophet/forecaster.py in setup_dataframe(self, df, initialize_scales)
286 df['cap_scaled'] = (df['cap'] - df['floor']) / self.y_scale
287
--> 288 df['t'] = (df['ds'] - self.start) / self.t_scale
289 if 'y' in df:
290 df['y_scaled'] = (df['y'] - df['floor']) / self.y_scale
/anaconda2/envs/mde/lib/python3.7/site-packages/pandas/core/ops/__init__.py in wrapper(left, right)
990 # test_dt64_series_add_intlike, which the index dispatching handles
991 # specifically.
--> 992 result = dispatch_to_index_op(op, left, right, pd.DatetimeIndex)
993 return construct_result(
994 left, result, index=left.index, name=res_name, dtype=result.dtype
/anaconda2/envs/mde/lib/python3.7/site-packages/pandas/core/ops/__init__.py in dispatch_to_index_op(op, left, right, index_class)
628 left_idx = left_idx._shallow_copy(freq=None)
629 try:
--> 630 result = op(left_idx, right)
631 except NullFrequencyError:
632 # DatetimeIndex and TimedeltaIndex with freq == None raise ValueError
/anaconda2/envs/mde/lib/python3.7/site-packages/pandas/core/indexes/datetimelike.py in __sub__(self, other)
521 def __sub__(self, other):
522 # dispatch to ExtensionArray implementation
--> 523 result = self._data.__sub__(maybe_unwrap_index(other))
524 return wrap_arithmetic_op(self, other, result)
525
/anaconda2/envs/mde/lib/python3.7/site-packages/pandas/core/arrays/datetimelike.py in __sub__(self, other)
1278 result = self._add_offset(-other)
1279 elif isinstance(other, (datetime, np.datetime64)):
-> 1280 result = self._sub_datetimelike_scalar(other)
1281 elif lib.is_integer(other):
1282 # This check must come after the check for np.timedelta64
/anaconda2/envs/mde/lib/python3.7/site-packages/pandas/core/arrays/datetimes.py in _sub_datetimelike_scalar(self, other)
856
857 i8 = self.asi8
--> 858 result = checked_add_with_arr(i8, -other.value, arr_mask=self._isnan)
859 result = self._maybe_mask_results(result)
860 return result.view("timedelta64[ns]")
/anaconda2/envs/mde/lib/python3.7/site-packages/pandas/core/algorithms.py in checked_add_with_arr(arr, b, arr_mask, b_mask)
1006
1007 if to_raise:
-> 1008 raise OverflowError("Overflow in int64 addition")
1009 return arr + b
1010
OverflowError: Overflow in int64 addition```
I understand that there's an issue with 'ds', but I am not sure whether there is something wring with the column's format or an open issue.
Does anyone have any idea how to fix this? I have checked some issues in github, but they haven't been of much help in this case.
Thanks

Answer 1

This is not an answer to fix the issue, but how to avoid the error.
I got the same error, and manage to get rid of the error when I reduce the number of data that is coming in OR when I reduce the horizon span of the forecast.
For example, I limit my training data to only start since 1825 meanwhile I have data from the year of 1700s. I also tried to limit my forecast days from 10 years forecast to only 1 year. Both managed to get rid of the error.
My guess this problem has something to do with how the ARIMA is implemented inside the Prophet itself which in some cases the number is just to huge to be managed by int64 and become overflow.