I have also thought of a methodology for type prediction which would be wholly empirically constructed, but it would require some level of statistical analysis which I am incapable of
. It's based on the construction of the Minnesota Multiphasic Personality Inventory, which was developed empirically due to issues where a questions designed to test for certain traits in practice elicited the wrong response (e.g. paranoid people were actually more like to answer "false" to the statement "I think most people would lie to get ahead" than a normal population). So it doesn't actually matter whether the statement is true or not, the only thing taken into account is what answer a specific group are more likely to put down.
Essentially it will be a questionnaire with lots (100-300) of seemingly arbitrary questions which don't appear to point to a specific type or dichotomy (Do you prefer red or blue? Do you consider yourself a good driver?) but are statistically correlated to a type (along the lines of Aquagraph's "everything is type related to some degree" thread).
The main advantage of this test is that it controls for confirmatory bias, which in my opinion is the main problem with current typing tests - if you have a reasonable grasp of the basics of socionics, you can easily guess the answers which would type you as whatever type you want to be. But with this test, the only way to "cheat" is if you know the statistical correlations of each random question to type. It would also control for exaggeration problems (e.g. IEIs responding to stereotype threat by answering "I am bloody awesome at sports thankyou very much! Intuitives can be good at sports too! I have trophies!"
)
The obvious problem with this test is constructing it - you would need a large and diverse population of people who are reliably typed to fill out massive questionnaires, then sort through all the questions to find the ones that are actually statistically significant, and analysing the hell out of a lot of data. But the construction process would actually be quite interesting since you'll actually end up with definitive and quantitative answers to "Is It Type-Related?".