twopirllc
commented
2 years ago Hello @tfgstudios,
Pandas TA is largely a Python implementation of TA Lib (and some few TradingView indicators) and thus the default mode for this Open Source implementation.
For the sake of brevity, I am only addressing ema.
This bug/feature sounds remarkably similar to Issue #420, TA Lib and it's Unstable Period as well as code and documentation of TA Lib's EMA for more details. I chose to implement TA_MA_CLASSIC computation.
Pandas TA currently has three options for an ema, help(ta.ema). You've have tried two of them (1 & 2).
- With TA Lib installed, it defaults to TA Lib.
- If
talib=Falseor TA Lib is not installed, it yields the same result as 1., as intended and you noted. - When
talib=Falseandpresma=Falseare arguments. (All three in the code and charts below).- This attempts to address:
-
Expected behavior EMAs values should not be affected by input values other than those mathematically required to calculate it
- If none of those suit your purposes, you are encouraged to include your own and submit it (preferably in numpy/numba). 😎 There are many implementations of "ema" beyond TA Lib's EMA. Unfortunately, I have not had enough time nor support to address them all.
_df = ta.df.ta.ticker("AA", timed=True)["Close"]
def bad_ema(src=None, length=None, n_samples=601):
if src is None:
src = pd.Series(np.linspace(40_000, 50_000, n_samples), dtype=float)
tal_ema = ta.ema(src, length=length, talib=True)
pta_presma_ema = ta.ema(src, length=length, talib=False)
pta_ema = ta.ema(src, length=length, talib=False, presma=False)
return pd.DataFrame({
"close": src,
f"tal_ema{length}": tal_ema,
f"pta_presma_ema{length}": pta_presma_ema,
f"pta_ema{length}": pta_ema
})
def cplot(df, last=None):
if isinstance(last, int):
df = df.iloc[-last:,:]
print(df.shape)
df.plot(figsize=(16,6), color=["black", "red", "orange", "green"] , grid=True)ma_length = 10
n_samples = 50
closedf = _df.iloc[:n_samples].copy()
df = bad_ema(closedf, length=ma_length, n_samples=n_samples)
cplot(df, last=None)
df.tail()
It is clearly evident that _ptaema10 (green line), in this case, adheres to "EMAs values should not be affected by input values other than those mathematically required to calculate it" expectation you desire. Which by one definition of ema relies on a minimum of two values.
Kind Regards, KJ
xucian
Which version are you running? The lastest version is on Github. Pip is for major releases. pandas-ta-0.3.14b0 (main) pandas-ta-0.3.65b0 (development)
Do you have TA Lib also installed in your environment? TA_Lib-0.4.24-cp38-cp38-win_amd64.whl
Did you upgrade? Did the upgrade resolve the issue? already at the latest version
Describe the bug EMA600's values are influenced by values that are even outside the last 600 observations. Shouldn't each value in an EMAX be obtained by the 'last X observations'? Why is it that even the ones before the last X observations (i.e. that aren't needed for EMA's calculation) affect its values? The difference isn't much, but if you're working on a 1m timeframe with an ema that has 256k data points (~6 months of data), the last value in that EMA would probably be very different from the last value of an EMA created from just the most recent 600 points (i.e. for simplicity, one with only 1 valid value -- the last one). Even if the difference wasn't much, my initial question holds. I'd like to disable any internal optimizations so that if EMA600 only needs 600 data points, it shouldn't care about- and be affected by what's before those 600 data points. How should I go about this without reinventing the wheel, i.e. creating a 600-len array for each individual step and create call ta.ema() multiple times?
To Reproduce
Gives:
Expected behavior EMAs values should not be affected by input values other than those mathematically required to calculate it
Additional context If we think about any indicator that requires 'the last x observations', is should be implemented this way: [Python-like pseudocode]
In a nutshell, we move from end to start, and get a mathematically accurate indicator value at each data point. I assume pandas goes from start to end or uses some arithmetic approximations, and this is happening regardless of whether I pass talib=False or True