The graduated rates are fitted using a function optimised using weighted least squares. Without performing further tests, how does one tell whether the graduated rates using this formula fits the experience better? The solution states that Cummulative Devs will now be zero, ChiSq test statistic will be greatly reduced (but won't be zero since the test looks at SSQ of the deviations), and serial correlation test will still be acceptable.
