17 write-ups · Rest in peace
The graveyard
Every strategy here, with why it died. Causes are explained in how we test; terms in the glossary.
Overfit 6
The winner doesn't survive an honest count of how many things we tried, or it's no better than scrambled data.
Diagnostics
144 signal tests
Two of 144 tests looked special; once we counted every try, they looked like luck.
Models
1,192 indicator combinations
The best of 1,192 mixes was no better than a search on scrambled data.
Trend
32 trend variants
32 trend rules were too alike; picking a “best” one was mostly noise.
Setups
24 “conditions align” setups
24 “wait for conditions to line up” setups; none beat random timing.
Models
28 indicator families and pairs
The best of 28 indicator cookbooks couldn’t be told from luck.
Models
Machine-learning models
Three models retrained each year; none beat a version trained on noise.
Edge decayed 2
It worked in one period and failed a later window we fixed in advance.
Inconsistent 1
The sign or result flipped between the windows we fixed in advance, so we don't treat it as a real effect.
Not tradable 1
The effect may be real but we couldn't have captured it with realistic trading rules.
Works but not worth it 5
It does something real, but misses its bar or gives up the whole return.
Trend
Plain 200-day rule
Cut the worst drops, but not enough to meet its 2017–2021 bar. It nearly matched the same 23-fund basket's return.
Calendar
Turn of the month
Those four days were a bit stronger, but sitting out the rest of the month earned little.
Calendar
Turn of month plus Fed days
Combining two calendar ideas added trades, not a working strategy.
Volatility
Volatility-targeted SPY/VTI/VUG
Cut the drop and nearly matched the market, but missed a bar we had already set.
Models
A take-or-skip judge
A stand-in test said yes to almost every signal; the original comparison was not completed.
Passed its test, no edge 1
It cleared the pre-registered test it was written for, but did not beat simply holding the market on the same days. It has not been traded with real money.
Untestable 1
A curiosity we would rather mark untestable than invent a p-value for.
Trial counts and rules are on each page. Based on public daily price data (see about).