I wonder if adjusting the factor weights or adding value-oriented screens could help address the low value factor loading. This framework may be worth clicking through to see how equal weighting impacts portfolio outcomes. Check this out, it relates! https://fungalstockecosystem.substack.com/p/i-got-a-25-return-in-a-backtest-then
Yeah, I think that's exactly it. To "perfectly" balance the risk exposures, you have to overweight the more stable low-turnover factors like value and quality vs. the faster factors like momentum, growth and sentiment.That's what I did back in my pre-ML days. Nowadays, I use a simple equal-weight anchor and let the ML algo overweight exposures implicitely based on market regime. Maybe in a future post I can repeat the Regression analysis but on simulated long-term returns and 3-year rolling periods to see if, when and how the ML algo shifts focus.
That would be a fascinating follow-up. The equal-weight anchor gives you a stable baseline, while the regime model can earn the right to deviate from it rather than constantly chasing whichever factor recently performed best.
The three-year rolling analysis would be especially useful because it could show whether the ML shifts are persistent, cyclical, or mostly noise. I’m working toward a similar idea with my allocator: keep the portfolio construction simple, then let evidence determine which factor families deserve more influence under different conditions.
I wonder if adjusting the factor weights or adding value-oriented screens could help address the low value factor loading. This framework may be worth clicking through to see how equal weighting impacts portfolio outcomes. Check this out, it relates! https://fungalstockecosystem.substack.com/p/i-got-a-25-return-in-a-backtest-then
Yeah, I think that's exactly it. To "perfectly" balance the risk exposures, you have to overweight the more stable low-turnover factors like value and quality vs. the faster factors like momentum, growth and sentiment.That's what I did back in my pre-ML days. Nowadays, I use a simple equal-weight anchor and let the ML algo overweight exposures implicitely based on market regime. Maybe in a future post I can repeat the Regression analysis but on simulated long-term returns and 3-year rolling periods to see if, when and how the ML algo shifts focus.
That would be a fascinating follow-up. The equal-weight anchor gives you a stable baseline, while the regime model can earn the right to deviate from it rather than constantly chasing whichever factor recently performed best.
The three-year rolling analysis would be especially useful because it could show whether the ML shifts are persistent, cyclical, or mostly noise. I’m working toward a similar idea with my allocator: keep the portfolio construction simple, then let evidence determine which factor families deserve more influence under different conditions.
I’d definitely read that post.
thanks for your interest in the topic.