getting error "ValueError: math domain error"

[razieh8051]

I am getting an error about the log function false input when using the sf.fit(df) for AutoARIMA. I checked my inputs of 'unique_id', 'ds', 'y' for df which are in correct format:
unique_id int64
ds datetime64[ns]
y float64
and there is no NaN or negative numbers that could cause this. Any ideas what could be wrong?

File ~\AppData\Local\Programs\Python\Python310\lib\site-packages\statsforecast\arima.py:1487, in Arima(x, order, seasonal, xreg, include_mean, include_drift, include_constant, blambda, biasadj, method, model, **kwargs)
   1485 nstar = n - tmp["arma"][5] - tmp["arma"][6] * tmp["arma"][4]
   1486 tmp["aicc"] = tmp["aic"] + 2 * npar * (nstar / (nstar - npar - 1) - 1)
-> 1487 tmp["bic"] = tmp["aic"] + npar * (math.log(nstar) - 2)
   1488 tmp["xreg"] = xreg
   1489 tmp["lambda"] = blambda

ValueError: math domain error

Update:
Here, in fm[i, i_model] = new_model.fit(y=y, X=X) of this line, y is supposed to be (n,) or (1,)? I figured in my input going into the loop it is (1,) and probably this what makes n negative in this line and then this math.log line cannot proceed.

ga in this line has the shape as (n, 1) <class 'numpy.recarray'> in both my input and your sample but when enters the loop for i, grp in enumerate(self): in the sample provided grp becomes (n, ) but in my input it becomes (1,); not sure why this is the case.

[tirthshah17]

Hello @razieh8051! I am getting similar error while running the fit method of statsforecast library. I checked my data and it does have any null values nor any negative values. Did you get any solution to this? If yes, then can you please share it?

[jmoralez]

Hey @razieh8051, sorry for the late reply. I've investigated this and there's a bug that only raises an error when there's a single sample, but it's wrong in general. I've opened #690 to track it. In the meantime you can define a fallback model in the statsforecast constructor that will be used when this error happens with arima or you can also drop series with only a single point.

Please let us know if you have further doubts.