Informer struggling to learn

I am experimenting with using the informer to forecast stock prices, and when I start training, it only ever gets as far as generating a noisy output relatively close to the target, not even a linear approximation:
I have experimented with context lengths from 8000 to 1000, lots of different learning rates, normalizing the data, and increasing the d_model and feed-forward layer dimensions, but nothing seems to improve it beyond the above picture. Any advice would be much appreciated!