Do LB6's stop-losses pay for themselves?
FRAME
- Subject: LB6 Config B, 6 BTC engines, each long-only or short-only (3 long, 3 short); 337 simulated trades, 2025-04-23 to 2026-08-22. Each engine has a fixed stop-loss: 3.5% to 8.0% from entry.
- Why it matters: 19 trades exited by stop-loss account for -85.9 of -186.6 points of losing trades (46%). Stops are an exit rule, which pak Yon's Config B revision itself changed.
- Prior work: lb6-opposite-weaker (SURVIVED), lb6-skip-opposite (FALSIFIED); neither touches exits.
- Current best: Config B as is: total +239.4%, mark-to-market max drawdown 8.6 pts, Calmar 27.99.
- Scoreboard: pak Yon judges by drawdown and Calmar: frozen-core3 was killed on drawdown budget, live runs a 15% drawdown kill switch, the 34-sleeve roster was picked by Calmar. [not confirmed for LB6]
- Question: stops take nearly half of all losses. Do they prevent larger losses, or cut trades that would have recovered by the engine's own exit?
- Scope: stop-loss only; entries, native exits and take-profits fixed. Research question, not a proposal to trade without stops. In-sample; decisive test = data after 2026-08-26.
- Terms: replay =
trading-infraoracledocs/lb6_tweaked/oracle/replay_lb6_tweaked.py(commit53c6993^) on inputs frommain; "no stop" = every engine's stop set so it never triggers, take-profit kept. Mark-to-market drawdown = deepest fall of the hourly portfolio value (closed net % plus each open position's % move at each Binance 1h close).
BET
Removing the stop-losses makes Config B more efficient.
- Adjective: more efficient
- why: the question is whether stops cost more profit than the risk they remove; efficiency weighs both.
- Measured by: Calmar = total net % / mark-to-market max drawdown, both variants computed the same way; higher = more efficient.
- why: drawdown from closed trades alone would hide the open losses a stop exists to cap; mark-to-market counts them.
- Because: stop exits carry 46% of all losses.
- Falsified if: Calmar(no stop) <= Calmar(Config B). Locked in
docs/bets/2026-11-17-lb6-no-stop.md, commit952f573, written before any replay ran.
KILL
Both variants replayed with the oracle; only the stop differs. Entries are identical (337 matched trades).
| test | falsified if | result | status |
|---|---|---|---|
| Calmar, mark-to-market | no stop <= 27.99 | 11.85 | FAIL |
Also observed, not deciding:
| variant | trades | sl_exits | total | mtm_max_dd | mtm_dd_trough | closed_max_dd | calmar_mtm | worst_trade |
|---|---|---|---|---|---|---|---|---|
| Config B | 337 | 19 | +239.4 | 8.6 | 2025-12-08 | 9.7 | 27.99 | -8.10 |
| no stop | 337 | 0 | +242.5 | 20.5 | 2026-08-25 | 24.4 | 11.85 | -23.53 |
Per engine:
| engine_id | stop_pct | n | sl_exits | total_stop | total_nostop | change |
|---|---|---|---|---|---|---|
| FOCUSED_TOP_SHARPE | 4.0 | 36 | 2 | +44.9 | +54.1 | +9.2 |
| deepdw_iv_M04 | 4.5 | 33 | 0 | +21.3 | +21.3 | +0.0 |
| deepput_strike_C14_G09 | 3.5 | 67 | 10 | +32.4 | +38.8 | +6.3 |
| dw_tail_C04 | 7.0 | 41 | 0 | +34.5 | +34.5 | +0.0 |
| iv_deepcall_0804_SL8 | 8.0 | 58 | 2 | +40.4 | +48.8 | +8.3 |
| netdw_skew_C01 | 5.0 | 102 | 5 | +66.0 | +45.2 | -20.8 |
Excluding the worst trade's window (from 2026-08-17 17:00 UTC, netdw_skew_C01 short, -23.53% without a stop): mark-to-market max drawdown 8.55 (2025-12-08) vs 13.24 (2026-06-05); the drawdown is still deeper without stops.
Prices: entries and native exits use the options-data spot; stop and take-profit checks and the hourly mark-to-market use Binance 1h bars (Binance Vision); only forced closes (data gaps, end of data) use merged_btc_hourly_FROZEN.csv, which ends 2026-07-14. One forced close falls after that: the worst trade above, closed at the end of data (2026-08-26 04:00 UTC), priced at the Binance 1h open for that hour (79,088.61, checked against the Binance Vision daily file). Last entry is 2026-08-22; the last exit is 2026-08-26. An independent re-implementation (subagent, no shared code) produced the same Calmar values.
The 19 stopped trades, with and without the stop (12 would have ended better without it, 1 of them by less than 0.1 pts; "saved" = with stop minus without, positive = the stop helped):
| engine_id | entry | net_pct_stop | net_pct_nostop | reason_nostop | saved |
|---|---|---|---|---|---|
| netdw_skew_C01 | 2025-05-08 | -5.10 | -5.41 | native_close | +0.31 |
| deepput_strike_C14_G09 | 2025-06-01 | -3.60 | +0.13 | native_close | -3.73 |
| netdw_skew_C01 | 2025-07-10 | -5.10 | -5.08 | native_close | -0.02 |
| deepput_strike_C14_G09 | 2025-10-10 | -3.60 | -8.93 | native_close | +5.33 |
| deepput_strike_C14_G09 | 2025-10-17 | -3.60 | -1.62 | native_close | -1.98 |
| deepput_strike_C14_G09 | 2025-11-13 | -3.60 | -6.26 | native_close | +2.66 |
| deepput_strike_C14_G09 | 2025-11-20 | -3.60 | -1.31 | native_close | -2.29 |
| FOCUSED_TOP_SHARPE | 2025-12-06 | -4.10 | +2.90 | TP | -7.00 |
| netdw_skew_C01 | 2026-01-10 | -5.10 | -4.39 | native_close | -0.71 |
| deepput_strike_C14_G09 | 2026-02-05 | -3.60 | +6.40 | TP | -10.00 |
| FOCUSED_TOP_SHARPE | 2026-02-13 | -4.10 | -1.87 | native_close | -2.23 |
| netdw_skew_C01 | 2026-03-03 | -5.10 | -7.89 | native_close | +2.79 |
| iv_deepcall_0804_SL8 | 2026-03-05 | -8.10 | -3.34 | native_close | -4.76 |
| deepput_strike_C14_G09 | 2026-03-27 | -3.60 | -3.23 | native_close | -0.37 |
| deepput_strike_C14_G09 | 2026-06-01 | -3.60 | -5.96 | native_close | +2.36 |
| deepput_strike_C14_G09 | 2026-06-03 | -3.60 | -6.31 | native_close | +2.71 |
| iv_deepcall_0804_SL8 | 2026-06-15 | -8.10 | -4.53 | native_close | -3.57 |
| deepput_strike_C14_G09 | 2026-07-23 | -3.60 | -2.56 | native_close | -1.04 |
| netdw_skew_C01 | 2026-08-17 | -5.10 | -23.53 | native_close | +18.43 |
UPDATE
- Status: FALSIFIED.
- What it means: without stops, total is +3.1 pts vs Config B and mark-to-market max drawdown is +11.9 pts; Calmar 27.99 -> 11.85.
- Next bets:
lb6-wider-stop(Does a wider stop on FOCUSED_TOP_SHARPE, deepput_strike_C14_G09 or iv_deepcall_0804_SL8 make Config B more efficient?) - Open: whether pak Yon's scoreboard for LB6 is Calmar; behaviour after 2026-08-26.