ML Model #2 Is Probably Done
When I launched on March 2nd I had a set of models I had spent a year tuning and backtesting, and one more I was still working on. I added it anyway. It cost me nothing but a column on the dashboard.
I did not hide what it was. The dashboard has labeled it "a portfolio based on an experimental model" since day one, and I called it exactly that in the Week 2 and Week 3 recaps. Unlike the others it is a single model, with no blend or ensemble behind it.
Then it won Iteration 1 at +20.33%, best of the four, so I kept it. If I am honest, the reasoning was "it is winning." That is not a reason. That is a two-month sample overruling a decision I had told myself I would make on the backtests.
What I Think Happened
| Iteration | Tech weight | ML Model #2 | vs S&P 500 |
|---|---|---|---|
| 1 (Mar 2 to Apr 30) | 10 of 15 | +20.33% | +16.64 pp |
| 2 (Apr 30 to Jun 30) | 9 of 15 | +32.73% | +28.47 pp |
| 3 (Jun 30 to Aug 31) | 14 of 15 | -16.15% | -19.67 pp |
| 4 (Aug 31 to Sep 8) | 3 of 15 | -4.44% | -4.38 pp |
For two iterations this was an AI and semiconductor bet wearing a machine learning model's clothes. That trade went vertical and the model looked brilliant. On June 30 it went to 14 of 15 in tech, doubling down almost exactly at the top, and lost 16% while the index gained 3.5%.
Then on August 31 it rotated nearly all the way out, down to 3 of 15, spread across nine sectors. Clean slate, no AI exposure left to blame. It is still losing, it is the only one of the four that is, and two of its fifteen picks are up.
That is the tell. A model with a real edge does not need one sector to work. This one caught a single enormous trend, rode it, doubled down too late, and has not found anything since.
What Happens Next
It stays until the October 29 rebalance. I am not killing it mid-iteration; the point of fixed 42-day windows and publishing picks in advance is that I do not get to change the rules once I can see how it is going.
But I would not bet on it. Since launch: Ensemble +38.68%, Growth Evaluator +34.76%, ML Model #2 +30.86%, S&P 500 +12.19%. All still well ahead of the index, but the model that led this experiment all spring is now last of the group.
The lesson is not a new one, which is the annoying part. Nothing goes live until it clears my full backtesting matrix: multiple windows, multiple regimes, a threshold written down before I look at the results. That process exists so a good-looking two months cannot talk me into anything. In March I skipped it because adding one more portfolio looked free.
It was not free.
See you at the October 29 rebalance.