Sparks on forecasting models

Just finished reading the set of papers titled Developments in Forecast Combination and Portfolio Choice, since it's due tomorrow. (I'm a bit lazy currently><)  Some of points I came to:

1. The difficulty of calculating downside standard deviation of a portfolio: Hmm...I've never realized such issue, maybe because I use absolute deviation directly... gonna check the model file later.
2. If buying on margin, downside SD is not enough to control the risk, VAR here should be brought in as an constraint to possible margin calls.
3. Unanimous vote vs. majority vote: why the former leads to a better result in selection? Should I use?
4. The idea of using accounting variables and nonaccounting varialbes are inspiring. I decide to use them as part of the inner personalities instead of as the front-end screening module as designed before.
5. Most trading system described still haven't address the issue of transaction costs. Moreover, our in-develop system has advantages on continous learning, portfolio-oriented (instead of single stock), and providing clear trading instructions.
6. The philosphy behind a model is much more critical than the algorithm used to implement it.