TA-Lib's STDDEV returning population and not sample.

[WitchGod]

Code first:

import pandas as pd
import talib as ta
import numpy as np

df = pd.DataFrame(np.random.randn(11), columns=['Values'])
df["pandas-std"] = pd.rolling_std(df["Values"], window=10)
df["talib-std"] = ta.STDDEV(df["Values"].values, timeperiod=10)

print df

Above gives:

      Values  pandas-std  talib-std
0   0.684808         NaN        NaN
1   0.613100         NaN        NaN
2  -0.016941         NaN        NaN
3  -0.519221         NaN        NaN
4   2.147814         NaN        NaN
5  -0.393357         NaN        NaN
6   0.123275         NaN        NaN
7   0.707619         NaN        NaN
8   2.157508         NaN        NaN
9  -0.138655    0.954126   0.905163
10 -0.028071    0.968347   0.918655

Using independent calculations, I come to the conclusion that TA-Lib calculates standard deviation based on population (divided by n) rather than by sample (divided by n-1).

Can an option be possibly added to STDDEV such that it's possible to switch between sample and population computations? The former is what's needed for calculating historical volatility.

[mrjbq7]

That appears to be the case, you can get ta.STDDEV's answer by doing a pd.rolling_std(..., ddof=0). Unfortunately, it appears the underlying C library that I am wrapping doesn't have the ability to parameterize that.

Is it not sufficient to be able to use pandas for this? Would you recommend trying to workaround the underlying library by adding that value myself?

Thanks,

[WitchGod]

No issues using pandas, really. Just wondering if this issue persists across other talib functions that uses observation periods, if any.

In any case, since historical volatility is a one-time computation and the other talib functions like BBANDS work well, I think using pandas is the best way to go for this.

The issue can be closed now. Thanks.

[asdf8601]

I've seen more or less the same issue so I wrote a snippet that hopefully can clarify how to solve this issue by making the nbdev parameter equivalent to the ddof:

>>> import talib
>>> import numpy as np
>>> import pandas as pd

>>> arr = np.array([1, 2, 3, 4], dtype=float)

>>> win = 3

>>> # mathematically both are equivalent  (proof of concept)

>>> np.std(arr[:win], ddof=1)
1.0

>>> np.std(arr[:win], ddof=0) * np.sqrt(win / (win - 1))
0.9999999999999999

>>> # using `nbdev` parameter

>>> talib.STDDEV(arr, win) * np.sqrt(win / (win - 1))
array([nan, nan,  1.,  1.])

>>> talib.STDDEV(arr, win, np.sqrt(win / (win - 1)))  # use nbdev parameter
array([nan, nan,  1.,  1.])

>>> # check with pandas

>>> pd.Series(arr).rolling(win).std(ddof=1).values
array([nan, nan,  1.,  1.])