PARALLAX-3
20d · spot · 36 products · coinbase · epoch 1
failed a gate; retained so the failure stays visible
Cross-sectional momentum over 20 days, skipping the most recent 2, rebalanced every 20 days, long the top 5 names by inverse-volatility weight at a 35% volatility target. Out of sample it returns -3.2% at a Sharpe of 0.11, against 10.0% for buying and holding BTC over the same window. Its in-sample Sharpe of 0.47 deflates to 0.00 once the 630 configurations searched are charged against it.
equity, in-sample then out
OOS sharpe
0.11
in-sample 0.47 → deflated 0.00
PBO
0.50
below the reject line
OOS return
−3.2%
max drawdown
-44.0%
the gauntlet
01
lookahead
02
survivorship
03
overfit
04
fills
05
regime
06
costs
This configuration held MPL-USD, BADGER-USD — products Coinbase has since delisted. They were traded all the way down and marked out at the last price they printed, not removed from the record.
This agent is retained on the board with its failure intact. It is not deployable and never will be — but deleting it would make the roster look like a process that produces winners, when it is a process that produces almost entirely losers and reports that honestly.
the file
forward window · never used for selection
The last 90 days were held back from the search entirely. Whatever this agent did here, it did without having been chosen for it.