Forecast vs. realized
FVR plots are useful for visualizing how well a model fits the data as the strength of the model forecast changes. Plots made on training data should show a roughly linear relationship, with slope=1. Plots made on test data usually show a weaker relationship, as the effects of overfitting are exposed.
Judging model quality is as much art as science. Of course we use metrics such as
R^2 in situations where we need to compare models automatically, such as during cross-validation. But the FVR plot shows so much more information. For example, for some datasets I don’t really care if my tiny forecasts near the origin are correct or not, since I won’t be trading those. I do care a lot about my forecasts far from the origin, as that is where I can overcome transaction costs and make money. I prefer when my large forecasts appear conservative. And I want my model to make lots of large forecasts!
Financial datasets tend to be large and extremely noisy. Raw scatterplots are useless, as the dots fill the plot area, and the clutter masks the structure. So, I prefer making binned scatterplots, where each dot represents the weighted average location of points in the bin. Here are examples from my handcrafted model on dataset-1 with varying numbers of bins. You can see how using fewer dots brings out the structure in the scatterplots: