macrocephalopod

@macrocephalopod may have an idea as well

Always continuous!

Assuming you are trying for a predictive regression, that is. Compressing to 0/1 is throwing away information, plus it gives you many more opportunities to overfit as you will be tempted to choose the threshold to get a good result.

Only reasons for using dummy variables are (a) you have a categorical (unordered) variable that you need to encode (b) the feature is essentially two-class anyway (e.g. very bimodal distribution) and you don’t think the variations around the peaks add any information.

Discretising continuous variables can also lead to jumps in your forecast if the underlying variable meanders around the threshold over time, which lead to undesirable turnover in your strategy (which costs you money)

Ah I understand now!! How about output? Is it generally to try to model a continuous variable (futureRet), or to a logit regression to model odds of a category (TopDecileRet)? If Y is cont, confused on how to trade a model, do you adjust your exposure as expctd y changs?

I exclusively try to predict continuous variables, generally either (excess) return over some period or (excess) return scaled by a volatility forecast

By “excess return” I mean return neutralised to any factors that you plan to hedge in your portfolio (eg market, sector or country exposure)