How would you best describe a statistical learning method that shows significant variation in results between multiple training datasets?

A statistical learning method that exhibits significant variation in results across different training datasets suggests that the method is highly sensitive to the specific data it is trained on. This characteristic is associated with high flexibility. Flexible models can fit a broad range of data patterns, including noise, leading to different outcomes based on the training set. This capability allows them to adapt extensively to the idiosyncrasies of each dataset, which can be particularly useful in capturing complex relationships but also carries the risk of overfitting.

While flexibility is advantageous in some contexts, it can result in a lack of consistency when applied to diverse datasets, as the model may latch onto irrelevant patterns that do not generalize well. In contrast, high bias would imply that the method consistently misses the underlying patterns, yielding similar results regardless of the dataset, while low variance would indicate more stability in predictions across different datasets. Saying a method is not suitable for predictive modeling overlooks the nuances of its flexibility; it may still perform well in certain scenarios but can be unreliable in terms of generalizing across varied datasets. Therefore, the defining feature of significant variation in performance across training datasets is the model's high flexibility.