Statsmodel using ARMA

Question

A bit new here but trying to get a statsmodel ARMA prediction tool to work. I've imported some stock data from Yahoo and gotten the ARMA to give me fitting parameters. However when I use the predict code all I receive is a list of errors that I don't seem to be able to figure out. Not quite sure what I'm doing wrong here:

import pandas
import statsmodels.tsa.api as tsa
from pandas.io.data import DataReader

start = pandas.datetime(2013,1,1)
end = pandas.datetime.today()

data = DataReader('GOOG','yahoo')
arma =tsa.ARMA(data['Close'], order =(2,2))
results= arma.fit()
results.predict(start=start,end=end)

The errors are:

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
C:\Windows\system32\<ipython-input-84-25a9b6bc631d> in <module>()
     13 results= arma.fit()
     14 results.summary()
---> 15 results.predict(start=start,end=end)

D:\Python27\lib\site-packages\statsmodels-0.5.0-py2.7.egg\statsmodels\base\wrapp
er.pyc in wrapper(self, *args, **kwargs)
     88         results = object.__getattribute__(self, '_results')
     89         data = results.model.data
---> 90         return data.wrap_output(func(results, *args, **kwargs), how)
     91
     92     argspec = inspect.getargspec(func)

D:\Python27\lib\site-packages\statsmodels-0.5.0-py2.7.egg\statsmodels\tsa\arima_
model.pyc in predict(self, start, end, exog, dynamic)
   1265
   1266         """
-> 1267         return self.model.predict(self.params, start, end, exog, dynamic
)
   1268
   1269     def forecast(self, steps=1, exog=None, alpha=.05):

D:\Python27\lib\site-packages\statsmodels-0.5.0-py2.7.egg\statsmodels\tsa\arima_
model.pyc in predict(self, params, start, end, exog, dynamic)
    497
    498         # will return an index of a date

--> 499         start = self._get_predict_start(start, dynamic)
    500         end, out_of_sample = self._get_predict_end(end, dynamic)
    501         if out_of_sample and (exog is None and self.k_exog > 0):

D:\Python27\lib\site-packages\statsmodels-0.5.0-py2.7.egg\statsmodels\tsa\arima_
model.pyc in _get_predict_start(self, start, dynamic)
    404             #elif 'mle' not in method or dynamic: # should be on a date

    405             start = _validate(start, k_ar, k_diff, self.data.dates,
--> 406                               method)
    407             start = super(ARMA, self)._get_predict_start(start)
    408         _check_arima_start(start, k_ar, k_diff, method, dynamic)

D:\Python27\lib\site-packages\statsmodels-0.5.0-py2.7.egg\statsmodels\tsa\arima_
model.pyc in _validate(start, k_ar, k_diff, dates, method)
    160     if isinstance(start, (basestring, datetime)):
    161         start_date = start
--> 162         start = _index_date(start, dates)
    163         start -= k_diff
    164     if 'mle' not in method and start < k_ar - k_diff:

D:\Python27\lib\site-packages\statsmodels-0.5.0-py2.7.egg\statsmodels\tsa\base\d
atetools.pyc in _index_date(date, dates)
     37         freq = _infer_freq(dates)
     38         # we can start prediction at the end of endog

---> 39         if _idx_from_dates(dates[-1], date, freq) == 1:
     40             return len(dates)
     41

D:\Python27\lib\site-packages\statsmodels-0.5.0-py2.7.egg\statsmodels\tsa\base\d
atetools.pyc in _idx_from_dates(d1, d2, freq)
     70         from pandas import DatetimeIndex
     71         return len(DatetimeIndex(start=d1, end=d2,
---> 72                                  freq = _freq_to_pandas[freq])) - 1
     73     except ImportError, err:
     74         from pandas import DateRange

D:\Python27\lib\site-packages\statsmodels-0.5.0-py2.7.egg\statsmodels\tsa\base\d
atetools.pyc in __getitem__(self, key)
     11         # being lazy, don't want to replace dictionary below

     12         def __getitem__(self, key):
---> 13             return get_offset(key)
     14     _freq_to_pandas = _freq_to_pandas_class()
     15 except ImportError, err:

D:\Python27\lib\site-packages\pandas\tseries\frequencies.pyc in get_offset(name)

    484     """
    485     if name not in _dont_uppercase:
--> 486         name = name.upper()
    487
    488         if name in _rule_aliases:

AttributeError: 'NoneType' object has no attribute 'upper'

Answer 1

When I run your code, I get:

"ValueError: There is no frequency for these dates and date 2013-01-01 00:00:00 is not in dates index. Try giving a date that is in the dates index or use an integer"

Since trading dates are happen at uneven frequency (holidays and weekends), the model is not smart enough to know the correct frequency for calculations.

If you replace the dates with their integer location in the index, then you get your predictions. Then you can simply put the original index back on the results.

prediction = results.predict(start=0, end=len(data) - 1)
prediction.index = data.index
print(prediction)

2010-01-04    689.507451
2010-01-05    627.085986
2010-01-06    624.256331
2010-01-07    608.133481
...
2017-05-09    933.700555
2017-05-10    931.290023
2017-05-11    927.781427
2017-05-12    929.661014

As an aside, you may want to run a model like this on the daily returns rather than on the raw prices. Running it on the raw prices isn't going to capture momentum and mean reversion like you probably think it would. Your model is being built off the absolute values of the prices, not on the change in prices, momentum, moving average, etc. other factors you probably want to be using. The predictions you're creating will look pretty good because they're only predicting one step ahead, so it doesn't capture the compounding error. This confuses a lot of people. The errors will look small relative to the absolute value of the stock price, but the model won't be very predictive.

I'd suggest reading through this walkthrough for a starter:

http://www.johnwittenauer.net/a-simple-time-series-analysis-of-the-sp-500-index/

Answer 2

Looks like a bug to me. I'll look into it.

https://github.com/statsmodels/statsmodels/issues/712

Edit: As a workaround, you can just drop the DatetimeIndex from the DataFrame and pass it the numpy array. It makes prediction a little trickier date-wise, but it's already pretty tricky to use dates for prediction when there is no frequency, so just having the starting and ending dates is essentially meaningless.

import pandas
import statsmodels.tsa.api as tsa
from pandas.io.data import DataReader
import pandas

data = DataReader('GOOG','yahoo')
dates = data.index

# start at a date on the index
start = dates.get_loc(pandas.datetools.parse("1-2-2013"))
end = start + 30 # "steps"

# NOTE THE .values
arma =tsa.ARMA(data['Close'].values, order =(2,2))
results= arma.fit()
results.predict(start, end)