Since I literally cannot help myself, I continued testing, and, found a flaw with the directional accuracy metric I was using, and added in some additional metrics for existing trend capturing for forecast selection.
I ended up with:
81% daily directional accuracy (returns a positive return 81% of days at the end of the day), with a 23% ROI over 27 days of historical trades with a 10-winner model.
If I reduce it to 3 winners, it jumps to 37% ROI over 27 days, but drops daily directional accuracy to 71% for the same time period. So, it has more days where it fails to return, but returns more overall.
Going to try a 3-winner run for a month and then a 10-winner run for a month if it doesn’t maintain the expected daily ROI. Either way, 71% and 81% daily directional accuracy is substantial. Finally.
I also built a pretty little graph that will show Tetsuo’s balance over time:
https://tetsuo.silogroup.org/performance_graph.html
I guess it’s time to shut up and let it run.
One mistake I was making with this was looking purely at individual forecast directional accuracy instead of grabbing the winners for each day being tested and calculating the total ROI, and, negative or positive, attributing a “win” or a “loss”. So, technically a return of 0.00001 is a “win” and counts towards directional accuracy in this new model, but, that’s alot better than a negative.
Overall this is performing well, but, if this ends up not being what the sims run, which has happened before, I may be nearing my limit of being able to improve it without swapping out the underlying time series forecasting engine.
As of 2024-11-25, it has 1030.98 for its balance. Let’s see where it’s at on Christmas morning.