Just saw this post.

My question is quite simple. Some businesses have highly cyclical ROE. UK housebuilders just now for example, ROE is extremely low; in boom times, it's very high.
How are you calculating ROE? Are you focusing on non-cylical sectors (medical devices!) or are you making a trend adjustment of some kind for cyclicals?

I love the five year average ROE figure at the FT global screener. They also have five year average ROA. But I don't have a way to backtest using that info, so I have to just trust that it's a great idea.

For backtests, when I use ROE I am usually looking at a fairly broad slate, not just a few firms, so the outliers (bad companies with just one good year) don't matter that much to the average, so I have generally just used the one-year ROE figure from Value Line because that's what I have. Which is I believe updated only once a year, but still offers a surprising amount of value.

e.g., 1998-2025 inclusive, using a universe which is a Russell 1000 proxy (top 1000 by market cap among those with US domicile):
Top 50 by ROE: CAGR S&P + 4.6%
Bottom 50 by ROE: CAGR S&P - 3.9%
Gap: 8.5%/year, nothing to be sneezed at for something so simple.

This doesn't mean that the few firms with the very highest ROE will be the best, but it's a good test just to skew your odds. Never turn down a tailwind.

One of my favourite simple screens, based on the observation that another good "general purpose" metric is sales growth:
S&P 500 proxy universe (largest 500 in the VL database with US domicile)
Ensure 5-year sales growth per share is positive, as is ROE
ROE * sales growth top 15
1997-2025, this beat the S&P by 6.3%/year.
Narrow it down to the 10 of those closest to their 52 week highs and it rises to an advantage of 8.9%/year. In backtest!

Same sort is nice among the Nasdaq 100.
e.g., Nasdaq 100 proxy
price to 52 week high top 50%, just to crowdsource away from the potential losers
ROE * sales growth top 5 or 10.
e.g., top 5 1998-2025 backtests at 9.4%/year better than the S&P.
That screen was put together in 2019; in the six years post discovery to 2025, top 5 returned 9.9%/year more than the S&P, suggesting it wasn't just a fluke of overtuning. Top 10 beat by 5.9%/year out of sample.
Rather remarkably, a gambler's screen of top 3 in those six years returned 29%/year and beat the S&P in all six individual calendar years. That won't continue, but it seems like a good omen, especially as the stretch 2020-2025 covers quite a variety of market conditions.

I don't run quant screens any more myself because I generally don't invest in the US any more--I'm down to only one US stock and a few derivatives, no cash or fixed income in USD. But I still like the MI art form.

Jim