Therefore, even the single-test Type I error rates are likely to be far from a nominal value like 0.05.ģ. I strongly doubt that a Gaussian or t-distribution is the proper reference distribution. You are testing hypotheses about single observations of unknown distributions. The t-test was designed to test hypotheses about means. You need a multiple comparisons adjustment to those testsĢ. The likelihood of falsely declaring a return as "significant" will be very high. With N = 40 (say) and 2 years, that would be 400 tests. With an average of N firms per year in each quintile, you will be doing 5N tests per year. I know nothing about financial statistics, but I question this whole project as it is now described.ġ.
Input float id int year float(CI_quintile weight AR wAR)ħ458 1988 1.
0 kommentar(er)
