I agree that the bar with 0 shift may have a future-looking problem. The version of Neuro Trend that I have uses bars at shift 1 and 2. There is no 0 shift bar. Perhaps we refer to different versions? My version has no version number. inputs 15 and 16 are:
input = gmThreshold((iMA(NULL,PERIOD_H1,5,0,MODE_EMA,PRICE_CLOSE,h1shift)-iMA(NULL,PERIOD_H1,5,0,MODE_EMA,PRICE_CLOSE,h1shift+1))/(Point*10));
input = gmThreshold((iMA(NULL,PERIOD_H1,5,0,MODE_EMA,PRICE_CLOSE,h1shift+1)-iMA(NULL,PERIOD_H1,5,0,MODE_EMA,PRICE_CLOSE,h1shift+2))/(Point*10));
They do not refer to bar 0.
However, we are really discussing how many angels can fit on the head of a pin. Neuro trend is a NN that predicts the next few bars of an EMA. look at the attached image.
Two curves are digital filters, the third is neuro trend. Which do you prefer? I cannot choose either.
Although neuro trend has a small mse, that is not very useful. The mse is measured over all samples (bars). It will be very small during smooth periods, and be larger when the ema pivots. But that is precisely when accuracy is needed. Since the pivots occur less often than smooth samples, they will have small effect on the mse. In fact, a simple linear predictor will have slightly worse mse, but be just as good at the pivots. Why bother with a NN?