On this page
The short answer
Should Birch hold only volatile stocks, the ones whose prices jump around most in a normal day? No. Built cleanly, the filter raised headline returns by leaning on a few lottery-ticket winners. It did worse in the tougher checks, and no version passed our luck test.
Many of the biggest winners in our data were the restless kind, so the idea sounds natural. An earlier machine-learning study made it look promising too. With the filter, Birch kept 29.2% a year with its 5 best stocks removed, against 27.5% without it. Before trusting that, we rebuilt the filter properly and tried hard to break it.
How did we test it?
The earlier version had two flaws:
- The scores it relied on exist only from 2019, so the filter was quietly switched off for 2016–18.
- In some test universes, that is the lists of stocks a test runs on, stocks without a score passed the filter anyway: 43 of 391 trades from 2019 on.
The clean version:
- The idea: take a new position only in stocks that move a lot for their price. Everything else stays the same.
- Two ways of deciding what "moves a lot" means: against the model's own past entries, or against all stocks on that day. Each was tried at three levels of strictness.
- Nine stock lists: the full list, and the same list with the biggest 50 to 200 stocks removed.
- The checks: 6 start dates, three times the slippage, the 5 best stocks removed, and 2023 onwards on its own.
- A random comparison: Birch skipping the same share of its entries at random, 20 times.
- A luck adjustment: a deflated Sharpe ratio across all 179 versions we tried. Sharpe measures return for the risk taken. The deflated version also counts how many versions we tried, so a lucky winner among many counts for less.
What did we find?
Rejected.
- The filter raised headline growth, and raised the 2023-onwards figure by up to 13 points. It did so by concentrating the portfolio. The share of profit from the single best trade rose from 6–13% to 20–33%.
- The tougher measures fell. From 2023 onwards, with the 5 best stocks removed, the middle setting lost to plain Birch in 8 of 9 stock lists, by 7.3 points on average.
- Take the stock list where the old study looked best, with the 100 biggest removed. There the clean filter gave 23.9% with the 5 best stocks removed, and 13.4% from 2023. Plain Birch gave 28.0% and 22.2%. The earlier 30.5% and 28.2% came from the two flaws.
- No version passed the luck test. The best deflated Sharpe was 0.51; we need 0.95.
What failed?
- The original lead came from missing data. There was no edge behind it.
- Comparing with all stocks on the day tied with Birch on average. It won on both tough measures in only 3 of 9 stock lists, with no pattern.
- Counting wins: 10 of 54 settings beat Birch on both tough measures, about what chance alone gives.
- The random comparison: skipping entries at random already gave a median of 25.4% on the full stock list. At one start date the filters beat 35–100% of the random versions. That is real but small, and it did not hold steady.
- The machine-learning versions built around it passed no luck test either: deflated Sharpe at most 0.04 across 244 tries.
Restless stocks do raise the chance of catching a huge winner. They also make the result depend on catching it.
What are the limits?
- The price data holds only companies still listed today. That flatters every result here.
- This study ran before exchange-traded funds (ETFs) were removed from the stock list. A re-run is planned.