Will increasing training ratio of the train test split always increase accuracy?

Question

I'm following an ARIMA tutorial from: https://machinelearningmastery.com/arima-for-time-series-forecasting-with-python/

He trains 66% of his data which results in a forecast and error like this:

ARIMA Rolling RMSE:  89.0210557987182
ARIMA Rolling MSE:  7924.7483755184985
ARIMA Rolling MAE:  68.66932854820298
ARIMA Rolling R2 Score:  0.21686864845457443

I then increased the amount of training to 70% which resulted in this:

ARIMA Rolling RMSE:  95.76039068274046
ARIMA Rolling MSE:  9170.052423711088
ARIMA Rolling MAE:  76.8507968436898
ARIMA Rolling R2 Score:  0.02235751415003584

Graph wise, both plots look very similar which I would have expected, however looking at the metrics, all metrics are much worse for the 70:30 split. Especially the R2 metric which only achieves 2% accuracy compared to 21% of the slightly longer training.

I increased and decreased the split for larger values and the same thing happened.

Is there a reason why this may be?