You misunderstood my approach: I wouldn't cache the signals (which is of course impossible because of the virtually unlimited number of combinations), but the predictions. The predicted value only depends on time, because the signals themselves only depend on time, unless you try to optimize parameters on which they depend. Is that even possible? At least for me, this isn't desirable. If it is possible, you could implement a flag DEEPCACHE or something, to switch caching on.


So if I was to implement that cache, I would pass a timestamp - a bar number or wdate() - along with the signals into the R session. So one of the columns of my signals.csv would be that timestamp. I would put that aside before my neural net gets trained. When the training is finished, I would first add a new column with my predictions to the signals table, then add the timestamp column. Then I would export only the timestamp and the predictions column to a file. So I would get a timeseries with all the predictions which I would read out from zorro.

Maybe the term "cache" doesn't describe the approach. It's more some pre-calculated time series.