Glad to see there are people on ET who have actually done some work... yes - mv optimization is very sensitive to inputs. This is a result of an implied 100% confidence in those expected return and var/covar numbers. In reality, we are never 100% certain of our forecasts (be it based on historical time series analysis or some other methods). There are some good refinements to mean-variance analysis that introduces forecast risk into those expected returns and var/covars. They tend to produce more 'sane' asset allocations. Black-litterman is one well known formulation - but I personally don't like it because it requires yet another set of inputs that are difficult to calibrate.
Can't - the Austrian School doesn't contain much of testable hypothesis that can be used as constraints or objectives. The Austrian School is a collection of (interesting) qualitative statements that aren't sufficient to form a scientific (that is, falsifiable) theory. Flamewar starts in 3...2...1....