X8: signal 669: rewrite the protection block

[guleconr]

Signal 669 (dump hunter, short, experimental) is disabled, and its protection block was mine, from #1201. It had 34 chains, 82% of them built from columns the rest of the file does not use (PH_*, SQZ_CNT, UO, WILLR_480, raw percent-change columns), and no chain in it was ever measured on its own. This replaces the whole block.

The new block is 153 chains. Every clause is a cached boolean that already exists in the entry scope, every threshold on a 0-100 column is a multiple of five, no chain has four clauses, and the clause order is low timeframe to high.

Rig

position_adjustment_enable: false, grinding off, grind_mode_max_slots: 0, 20 slots, signal 669 added alone to the standing enabled set. Three populations: binance 2022 (bear), bybit 2026 and binance 2026. The profit-max cache is deleted before every run. Every number below is a real backtest, not an offline projection.

Result

block	trades	losses	tag P&L
raw, no protections	8,326	506	-16,593
the 34 chains being replaced	not re-measured
this block (153)	2,097	28	+99,131

Per population, against the last intermediate block so the last step is visible:

	137 chains	153 chains
binance 2022	708t / 7L / +39,088	682t / 7L / +36,064
bybit 2026	732t / 22L / +22,180	722t / 12L / +36,444
binance 2026	710t / 16L / +20,271	693t / 9L / +26,623

53 trades bought 17 losses and 22% more money.

How it was built

The backbone came from a measured round; everything after it is per-loss work. The last two passes read the remaining losses one at a time and wrote a chain for each at zero winner cost in its own population — 16 chains, 14 of them three-clause.

Binance 2022's twelve residual losses split into four scenes:

scene	losses	reading at entry
upper timeframes already spent	ALGO + FIL 10-13, SAND 07-26, SOL 11-22	rsi_3_4h 1.5 / 3.4, rsi_3_1h 0.5, stochrsi_k_1d 0.0, willr_14_4h -99.6
capitulation reversal	XRP + DOGE 01-24	rsi_3_change_pct_1d +2269% / +785% against a winner median of -9.65
the 4h has turned up	SOL + AAVE 01-31	change_pct_4h +1.3 / +0.8, 98th percentile
small change	AXS, FIL, ADA, AVAX	no common axis

The first scene is the signal's own blind spot: the Logic checks roc_288 > -10, so it already refuses to sell a 24h that has crashed — but nothing checked whether the 1h, 4h and 1d were spent. Those chains are the complement of a filter the signal already has.

What was measured and rejected

• Running the round a second time on its own blocked export (164 chains). Offline it takes binance 2022 to zero losses. Run for real it gives 369t / 2L / +13,515 against 137's 722t / 12L / +31,965: ten losses cost 343 winners, 34 per loss, against a break-even of 12.3, and half the money. Not shipped.
• Chains for the losses that do not separate. SOL and AAVE 01-31 sit in the middle of every distribution; the cheapest chain reaching either costs 176 to 199 winners. DOGE 01-24 costs 39 plus 54 more on another population, ADA 128, FIL 71. Those five losses are left in on purpose.
• Two candidates dropped for house style, not for their numbers: they needed rsi_3_lt_15 / rsi_3_lt_20, which are not cached in the entry scope, and two more needed non-round raw thresholds (roc_9 < -0.5, roc_9_15m > -2).

The 28 that remain

They are not fast squeezes. Most stay open longer than two days: the break is correct and the market walks over it. Four independent attempts to separate them at entry failed, which is what you would expect from a signal that fires at the break rather than after it — the winner and the loser are the same candle. 170 could be reduced to five losses because it fires after capitulation, when the context already describes the move.

short_entry_condition_669_enable stays False. ruff format and ruff check clean at the pinned version, py_compile clean, and the bare block the measurements ran on is byte-identical to the one in this diff.

🤖 Generated with Claude Code