Step 21 — Weekly live regime

What: replace the Step 19 manual regime input with the simplest causal market state that still looks live-usable. Why: remove oracle-style routing input before further paper-runner work.

Step 21 weekly live full return

9.34%

Step 19 manual benchmark

104.93%

Step 20 simple v2 full return

86.59%

Open position

flat / no open position

Detector definition

Benchmark asset: ETH/USDT.
State timeframe: weekly bars built from the same live-style ETH data feed.
Rule: bull if close > 40w MA, 12w MA > 40w MA, and 12w return > 12%; bear if the symmetric bearish condition holds; else sideways.
Causality guard: each hourly bar uses the last completed weekly state only, implemented as a 1-week shift before forwarding to the runner.
Router held constant: bull dip-buy 24/4/16, bear rally-fade 24/4/16 with dist200 filter, sideways no trade.

Full-sample comparison

Variant	Trades	Total return	Max DD	Win rate	Profit factor
Step 19 manual benchmark	115	104.93%	-13.67%	58.26%	1.79
Step 20 live-style v1	38	60.87%	-15.14%	63.16%	2.46
Step 20 simple v2	56	86.59%	-12.22%	58.93%	2.41
Step 21 weekly live	116	9.34%	-22.02%	54.31%	1.12

Observed state mix

Bull: 22.96%
Bear: 14.39%
Sideways: 62.65%
Latest state: bear at 2026-03-11T23:00:00Z

Honest read

This does remove the manual regime dependency. The state is causal, simple, and live-native enough to wire into the runner.
It is not competitive with the current benchmark: 9.34% vs 104.93% for Step 19, a gap of -95.60%.
It is also worse than the simpler hourly detector from Step 20: 9.34% vs 86.59%, a gap of -77.25%.
The weekly state traded too often for the quality delivered (116 trades) and carried materially deeper drawdown than Step 20 simple v2.
Conclusion: useful as a real live-native baseline, but not good enough to become the default routing state for continued paper-runner development.