Suppose I wish to estimate a logit model to predict Large losses when trading stocks, well I might say y_t = 1 if that trade was a large loss. But when I would then try to estimate a model to predict these large losses I would essentially just be modelling which stocks are most volatile. Ie. I would throw away both my best trades and my worst ones. Could I modify the logit model such that it avoided this ? - I was imagining adding a penalty to the log-likelihood contributions, such that a missclasification of a great trade was penalized more heavily than a mediocre one.
thoughts ?
thoughts ?