After getting back in the office, I start reviewing the results from the run that wasn’t complete before I left. I may have found that one of the four parameters I’m dealing with just function of one of the others, so it may be able to be eliminated or significantly reduced.
I’ve started another run in an adjacent neighboring parameter space to see if the pattern holds there as well. I’ve also enhanced my summarizer app to filter out results for parameters outside of certain ranges. I first added filtering so I could verify that using the values of the fourth parameter that would correspond to it just being a function of the another actually improved the overall results. It did, so that’s good.
Also, with that first filter in place a clear pattern emerged in the metrics and I found that only 2 of the 64 metrics were really useful (and maybe actually only one) at the second metric layer. A good number of the metrics were redundant, producing exactly identical results even with results averaged together from over a dozen different stocks.
I had noticed previously that the metrics that were coming out the best when using two layers of metrics were not the same as when using one layer of metrics, so I had suspected that different metrics might best at each layer. I’ll need to refactor some code to make that happen and I’ll likely need to go and do some runs with just one layer of metrics again, to try cut down that layer’s parameter space before adding the second layer back on and then trying to think up some new metrics to try at that layer.