Different results in ndiffs pmdarima (Time Series)

Question

I am analyzing a time series. It clearly has a trend and a seasonal component. When I do the adf root test I get a p-value of 0.98, meaning it's non stationary.

But when I do the ndiffs in pmdarima, Philippe Perron and Dickey Fuller returns a 0, when clearly has a trend. KPSS return 1, which seems more accurate. Happens the same when do the nsdiffs, when clearly is stationary.

What I am doing wrong? Why I get different results? Does this mean that doesn't have a trend?

from pmdarima.arima.utils import ndiffs

difs_adf = ndiffs(train_, test = "adf")
difs_kpss = ndiffs(train_, test = "kpss")
difs_pp = ndiffs(train_, test = "pp")
print(difs_adf , difs_kpss , difs_pp)

On the other hand, what means the "regression" parameter in adf test in stattools. Is "ct" used when the series has a trend? constant is synonim of heteroskedasticity?

from statsmodels.tsa.stattools import adfuller, kpss

adf = adfuller(train_, regression='ct', autolag='AIC')
print(f'ADF Statistic: {adf[0]}')
print(f'p-value: {adf[1]}')

Answer 1

Not sure if helps but I found this example where the approach is to take the max result of these tests and continue with it: https://notebook.community/tgsmith61591/pyramid/examples/stock_market_example

from pmdarima.arima import ndiffs

kpss_diffs = ndiffs(y_train, alpha=0.05, test='kpss', max_d=6)
adf_diffs = ndiffs(y_train, alpha=0.05, test='adf', max_d=6)
n_diffs = max(adf_diffs, kpss_diffs)

print(f"Estimated differencing term: {n_diffs}")