which mechanisms repeatedly fail;
which ideas only worked before 2020 and then broke;
which combinations reduce drawdown but destroy CAGR;
whether some rejected strategies might be useful as low-correlation portfolio components despite poor standalone CAGR;
and which research directions we should stop wasting time on.

I am not suggesting you publish 1,900 long strategy pages. A downloadable machine-readable rejects table would be ideal though. Even strategies with poor standalone CAGR can be useful for understanding which mechanisms fail consistently, which ones broke out of sample, and whether any low-CAGR ideas have unusually low correlation or drawdown characteristics that could still make them useful as portfolio components. A downloadable table would be much more useful than 1,900 separate pages.

two CSVs.

One row per strategy with fields such as strategy ID, universe, mechanism, rebalance frequency, holding count, ordered rule stages, timing rule, weighting, transaction-cost model, sample dates, OOS dates, CAGR, MDD, Sharpe, Sortino, Ulcer, volatility, beta, percent cash, turnover, worst rolling 1/3/5/10-year results, recovery time, publication date, and incubation end date.

Then a separate monthly-return file with:

strategy_id, month, monthly_return

For the rules, an ordered stage format would be especially useful—for example stage 1 filter/rank/keep, stage 2 filter/rank/keep, stage 3 final selection—rather than only one prose description.

Your explanation of the monthly execution timing is also useful. I use a different convention: signal on the third-to-last trading-day close and execute at the second-to-last trading-day close. I would be very interested in seeing whether your results materially change when you test that timing.

I am also interested in the rejects table. If you publish it, I would especially want to distinguish “negative full-history CAGR” from “positive full-history but negative 2020+ OOS CAGR,” since those tell rather different stories.

The parameter-testing explanation is helpful too. The mean-minus-one-SD calendar-year selection is sensible. The neighboring-parameter grid would be useful because it can identify whether a result lies on a broad plateau or is a narrow parameter spike.