ADX value from plugin vastly different from Tradingview, just me or did I mess up some entry?

Hi,

So first off this problem persists across multiple plugins and code. I love the pandas_ta library thus was hoping we could find a solution here.

First off here is the difference in values:

Here is my code:

def each_copy(ticker):
    ticker_data = pd.DataFrame.copy(df[df["ticker"]==ticker],deep=False)
    ticker_data.set_index('date',inplace=True)
    return ticker_data

test_data = each_copy("MRPL")
test_data["adx"] = ta.adx(high=test_data.high,low=test_data.low,close=test_data.close)["ADX_14"]
test_data["aDMP_14"] = ta.adx(high=test_data.high,low=test_data.low,close=test_data.close)["DMP_14"]
test_data["aDMN_14"] = ta.adx(high=test_data.high,low=test_data.low,close=test_data.close)["DMN_14"]
test_data.loc["2022-04-19 15:15:00+05:30":]

Data from the function ‘each copy’ is returned this way

After adding adx data

I hope this is enough, I did notice that the code on Tradingview is different from what we are using here. Could that be the reason? Does TV use some other formula for ADX?

I am not much of a coder thus relying on your help.

Please help and thank you.

[hakumaku]

The reason adx value is different is while TradingView handles the initial value differently, pands-ta does not.

I faced this issue recently, and I don’t know about the very detail of adx indicators, but I figured out where the difference comes from by looking at the code.

As you might know, rma depends on the previous value to get the current value. How to set the first value of rma seems to vary but TradingView uses the average of values.

The below is the code block of adx function. The version(master, dev) does not matter, because both do not seem to handle the first value.

def adx(
    high: Series, low: Series, close: Series,
    length: Int = None, lensig: Int = None, scalar: IntFloat = None,
    mamode: str = None, drift: Int = None,
    offset: Int = None, **kwargs: DictLike
) -> DataFrame:
    ...
    k = scalar / atr_
    dmp = k * ma(mamode, pos, length=length)
    dmn = k * ma(mamode, neg, length=length)

    dx = scalar * (dmp - dmn).abs() / (dmp + dmn)
    adx = ma(mamode, dx, length=lensig)
    ...

All of dmp, dmn and adx depend on ma("rma", ...) method.

If you delve into the function deeper, it basically uses pandas.DataFrame.ewm method.

It has adjustbool, default True parameter that you should pay attention to.

Setting it to True acts as the following: $y_t = \frac{x_t + (1 - \alpha)x_{t-1} + (1 - \alpha)^2 x_{t-2} + … + (1 - \alpha)^t x_0}{1 + (1 - \alpha) + (1 - \alpha)^2 + … + (1 - \alpha)^t}$

False which is what we want (to sync with TradingView): $\begin{split}\begin{split} y_0 &= x_0 \ y_t &= (1 - \alpha) y_{t-1} + \alpha x_t, \end{split}\end{split}$

It seems like it does not provide a way of handling $y_0 = x_0$ part, and I couldn’t find it either. (someone asked the question at stackoverflow)

To work around this… you should hack input values a bit to get desired result:

    # Set the initial value to sum() of 'length'.
    pos.iloc[length-1] = pos[:length].sum()
    pos[:length-1] = 0
    neg.iloc[length-1] = neg[:length].sum()
    neg[:length-1] = 0
    ...
    dmp = k * pos.ewm(alpha=alpha, adjust=False, min_periods=length).mean()
    dmn = k * neg.ewm(alpha=alpha, adjust=False, min_periods=length).mean()

adx is a bit more tricky. It requires length of both dmpand dmn values (at least for TradingView). If your data starts from 06-19, the first vaue of adx will appear at 07-16 which is 27 days later assuming the length is set to 14.

    ...
    # Set the initial value to sum() of 'length'.
    dx = dx.shift(-length)
    dx.iloc[length-1] = dx[:length].sum()
    dx[:length-1] = 0

    adx = dx.ewm(alpha=alpha, adjust=False, min_periods=length).mean()
    adx = adx.shift(length) # rollback

I’m not sure this is considered to be a bug or something, but I hope my explanation helps.

[anon2010]

How can I implement this fix? Swap out Hakumaku’s def adx method for the one in adx.py ?