stationarity and fractional differencing

This is a methodology question. I would like to make the data stationary but not transform it "too much" (information loss), before it is fit for statistical/ML purposes such as regression or PCA.

Typical procedure:

1. We inspect the raw time series $$x_t$$. Assume we see trends or if the ADF test fails, so I seek to transform the data.

2. We check that differencing $$x_{t+1} - x_t$$ results in stationary series. (Assume it does.)

3. If possible, we seek alternatives to $$x_{t+1} - x_t$$: the transformation may excessively remove memory from the series. That may help stationarity but at the price of weaker correlations to other series. We therefore use fractional differencing, say with d=0.5, and check that the series is more stationary than the initial series.

Let's apply this to the following series:

Conclude that it is stationary with $$95\%$$ confidence when using ADF implemented by adfuller() from Python's statsmodels package. The default parameter is $$12(n/100)^{1/4}$$ where $$n$$ is the series length. However, we see some (noisy) trend between 2011 and 2016. For the first half of the data the values are almost all negative. I conclude that the data doesn't give enough "labels" for model training.

I am therefore not sure whether the ADF test can be trusted. Would you feel comfortable using it for regression/PCA? Now compare with the fractional differencing with $$d=0.5$$:

As expected this makes the series look more stationary. However, it also decreases the correlation to other similar dataset series transformed in the same way. Average correlation decreases as I increase $$d$$, as can be expected.

I therefore need to choose the smallest $$d$$ such that I feel comfortable that the transform is stationary. How do you know when to stop? I am reluctant to use the ADF pvalue as threshold to automatically set this considering the issue above.