# The settings that won before 2019 were no better than random after it — retained record

> Retained reddit record from reddit:algotrading.

Canonical: https://highsignal.app/data/records/92dd36207555ecfb

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:algotrading)  published 2026-10-13 14:35 ingested 2026-10-14 02:44

[original source ↗](https://www.reddit.com/r/algotrading/comments/1x4ybpc/the_settings_that_won_before_2019_were_no_better/)

## retained text



```
A broad profitable parameter surface can coexist with almost no persistence in which settings win. I tested that distinction with 700 configurations of a monthly momentum and correlation allocation rule.

The base rule ranks 14 global asset classes on average 3, 6 and 12-month momentum, keeps the top 5, then holds the least-correlated 3 using the previous 252 trading days, with equal weights and a cash filter. I varied momentum lookbacks, shortlist size, holdings count and correlation window. Each configuration ran over 1987 to 2026.

I split the comparison into 1987 to 2018 and 2019 to 2026, ranked the configurations by CAGR within each window, and compared their ranks. The correlation was 0.048. The published settings moved from 3rd to 58th among the 700.

|Published settings|CAGR|Rank among configurations|
|:-|:-|:-|
|1987 to 2018|14.50%|3rd|
|2019 to 2026|16.36%|58th|

The return improved while the relative rank deteriorated. I can't use the higher later CAGR as evidence that choosing those settings worked. All 700 configurations exceeded 10% CAGR after 2019. The later period favoured the family broadly.

[The parameter sweep tests a family of settings; the split-period comparison checks whether their ranks persisted.](https://preview.redd.it/tznl5l6zw8vh1.png?width=2100&format=png&auto=webp&s=63dabef3f26e9957bb744e57b7db85427de4deb0)

I also inspected the 19 neighbours that differ from the published settings on exactly one axis. Their full-period CAGRs ranged from 10.6% to 15.0%. The median neighbour gave up 1.51 percentage points of annual compound return; the worst gave up 4.07. Those are losses relative to the published configuration, not negative portfolio returns.

My interpretation is narrower than "the parameters don't matter". I didn't find a collapse from moving a setting, but the split also didn't show that the best fitting-window rank carried forward. Ranking noise can select a winner even when the surrounding rules all benefit from the same market behaviour.

There are material limits. The grid is centred on the published values, so I chose the neighbourhood. Universe choice remains outside this parameter sweep. The claimed 2019 design date has no public timestamp; I can't establish a genuine holdout from that assertion. This split is a retrospective diagnostic, and a friendly later regime doesn't establish an edge.

When you tune a systematic rule, what do you hold back as a genuine out-of-sample window, and how do you stop yourself from looking at it early?
```



document fetched 2026-10-14 00:44 · canonical https://reddit.com/r/algotrading/comments/1x4ybpc/the_settings_that_won_before_2019_were_no_better
