I suspect the key distinction may the difference between extrapolation and interpolation. The outside view is far and away more reliable for interpolation: i.e. we have many events on both sides of the one we’re attempting to model. Extrapolation, where we have data only on one side of the event we’re trying to model, runs a much greater risk of failing in the outside view. The risk of failing in the inside view is no less in this case than with interpolation, but in this case the outside view is not as strong a predictor.
I suspect the key distinction may the difference between extrapolation and interpolation. The outside view is far and away more reliable for interpolation: i.e. we have many events on both sides of the one we’re attempting to model. Extrapolation, where we have data only on one side of the event we’re trying to model, runs a much greater risk of failing in the outside view. The risk of failing in the inside view is no less in this case than with interpolation, but in this case the outside view is not as strong a predictor.