# Stop-loss overlay for BTC funding mean reversion

**Date:** 2026-11-14  
**Strategy:** `brain/strategy/2026-11-14_btc_funding_mean_reversion`  
**Question:** Should the 48h time-exit be supplemented with a price stop? Does a
1-minute stop scan make sense?

---

## 1. Current exit rule

The live strategy has **no price stop**:

- enter when `funding z ≤ −1.0` and `ATR distance > 0.5` (long), or
  `funding z ≥ +1.0` and `ATR distance < −0.5` (short)
- hold exactly **48 hours**
- then flat for **24 hours** before the next entry

The worst single-hour loss in the official backtest was −15.9%; the worst
single trade in this study is −14.0%. A 30-minute crash is not avoided.

## 2. Method

- **Entries fixed.** We take the baseline signal's entries and only change the
  exit. This isolates the stop effect; it does not re-enter earlier after a
  stop.
- **1-minute simulation.** Stops trigger on Bybit-linear 1m lows/highs. If a
  1m bar opens beyond the stop, the fill is the open (gap); otherwise the stop
  price. **5 bps slippage** on stop fills, 3.5 bps fee per side.
- **Metrics.** 1m PnL → hourly PnL → geometric CAGR / Sharpe / MDD.
- **Split.** 2020-04→2024-11 back, 2024-11→2026-11 forward (the forward window
  was already inspected in earlier strategy work; it is validation, not a
  pristine holdout).
- **Variants.** No stop; fixed 3/5/8/10/15%; ATR 3/6/12/24 (ATR is the 14-day
  average 1h true range at entry).

Baseline check: the reconstruction gives CAGR 50.9% back / 46.2% forward and
Sharpe 1.45 / 1.48, matching the official 1m report (50.7% / 48.9%, 1.40 /
1.56). The geometric MDD (−37.7% / −26.6%) is larger than the report's
arithmetic `max_drawdown_percentage` (−24.7% / −23.2%); this study uses
geometric MDD, as required by the research contract.

## 3. Results

### Forward (2024-11 → 2026-11)

| Variant | CAGR | Sharpe | MDD | Worst trade | Stops | Stop rate | Whipsaw net | Premature |
|---|---:|---:|---:|---:|---:|---:|---:|---:|
| No stop | 46.2% | 1.48 | −26.6% | −8.8% | 0 | 0% | 0 | 0% |
| Fixed 3% | 35.6% | 1.30 | −19.1% | −3.1% | 36 | 24.3% | +0.166 | 58% |
| Fixed 5% | 49.0% | 1.58 | −23.5% | −5.1% | 9 | 6.1% | −0.030 | 56% |
| Fixed 8% | 46.6% | 1.49 | −26.5% | −8.1% | 3 | 2.0% | −0.005 | 33% |
| Fixed 10% | 46.2% | 1.48 | −26.6% | −8.8% | 0 | 0% | 0 | 0% |
| 3 ATR | 34.4% | 1.33 | −16.9% | −4.4% | 63 | 42.6% | +0.221 | 56% |
| 6 ATR | 41.1% | 1.39 | −25.9% | −8.7% | 24 | 16.2% | +0.073 | 63% |
| 12 ATR | 46.3% | 1.48 | −26.6% | −8.8% | 1 | 0.7% | −0.006 | 0% |

### Back (2020-04 → 2024-11)

| Variant | CAGR | Sharpe | MDD | Worst trade | Stops | Stop rate | Whipsaw net | Premature |
|---|---:|---:|---:|---:|---:|---:|---:|---:|
| No stop | 50.9% | 1.45 | −37.7% | −14.0% | 0 | 0% | 0 | 0% |
| Fixed 3% | 51.6% | 1.74 | −32.3% | −3.1% | 73 | 29.0% | +0.037 | 53% |
| Fixed 5% | 51.2% | 1.60 | −38.8% | −5.1% | 35 | 13.9% | +0.014 | 49% |
| Fixed 8% | 51.8% | 1.53 | −38.8% | −8.1% | 13 | 5.2% | −0.031 | 39% |
| Fixed 10% | 46.9% | 1.41 | −44.4% | −10.1% | 10 | 4.0% | +0.080 | 70% |
| 3 ATR | 42.2% | 1.49 | −41.2% | −7.6% | 109 | 43.3% | +0.381 | 63% |
| 6 ATR | 48.5% | 1.48 | −42.4% | −15.1% | 38 | 15.1% | +0.118 | 53% |
| 12 ATR | 46.7% | 1.37 | −41.7% | −19.0% | 4 | 1.6% | +0.060 | 75% |

**Whipsaw net** = sum of `time-exit PnL − stop PnL` over stopped trades.
Positive means the stop destroyed value (the trade would have recovered);
negative means the stop saved value. **Premature** = share of stopped trades
with a positive whipsaw.

## 4. Interpretation

### Tight stops hurt the edge

- Fixed 3% cuts forward CAGR from 46.2% to 35.6% and lowers Sharpe from 1.48 to
  1.30. 3 ATR is worse: 34.4% CAGR, 42.6% stop rate, +0.221 whipsaw.
- The mechanism is structural: the strategy buys fear / sells greed. A tight
  stop sells exactly when the crowd is most extreme, which is often just before
  the reversion. 53–63% of stops were premature.
- The MDD improvement (−26.6% → −19.1% / −16.9%) is real but bought with a
  large return sacrifice. It is not a good trade.

### A wide catastrophic stop is defensible

- **8% fixed**: CAGR 46.6% vs 46.2%, Sharpe 1.49 vs 1.48, MDD −26.5% vs
  −26.6%, worst trade −8.1% vs −8.8%. Nearly free tail insurance. Whipsaw is
  −0.005 (neutral).
- **5% fixed**: forward CAGR 49.0%, Sharpe 1.58, MDD −23.5%, worst trade −5.1%.
  Better forward metrics, but the backtest MDD is slightly worse (−38.8% vs
  −37.7%) and whipsaw is +0.014 back / −0.030 forward. Given the forward window
  is contaminated, treat the 5% improvement as a hypothesis, not a result.
- **10% / 15%**: trigger too rarely; when they do, 70–100% are premature, and
  they can worsen MDD (10% back MDD −44.4%, 15% −46.2%). Not useful.
- **ATR stops are inferior here.** ATR is a 1h range; 3 ATR is too tight for a
  48h hold, 6 ATR is worse than 5–8% fixed, and 12/24 ATR almost never trigger.

### The stop does not fix the drawdown

Backtest MDD stays −32% to −39% for every variant. The drawdown is a sequence
of losing trades (2021, early 2025), not one catastrophic trade. A stop caps the
single-trade tail; it does not cap the portfolio path. Position sizing / vol
targeting is the first line of defense.

### Does a 1-minute scan help?

Frequency is not the main lever. The rule and the level are. A 1-minute scan
with a 3% stop still loses to the 48h exit because the stop is inside the
normal noise band. If a stop is used, put it at the exchange as a wide
catastrophic order (5–8%), not as a tight per-minute bot loop.

## 5. Recommendation

1. **Do not add a tight stop** (3%, 3 ATR). It destroys the mean-reversion edge.
2. **Add an 8% fixed catastrophic stop** if a hard loss cap is required. It is
   the most neutral across both windows and caps the worst trade from −14.0% /
   −8.8% to −8.1%. Use 5% only if a tighter tail is worth the backtest MDD
   cost.
3. **Keep the 48h time exit** as the primary exit; the stop is a tail overlay.
4. **Re-run the official strategy** with the chosen stop, then paper trade. Do
   not optimize the stop level further on the 2024–2026 window.
5. **For the drawdown**, test vol targeting / position sizing separately; the
   stop is not a substitute.

## 6. Caveats

- **Forward is contaminated.** The 2024-11→2026-11 window was inspected during
  earlier strategy selection. It is validation, not a clean OOS.
- **Entries are fixed.** A stop would free capital earlier and could change
  re-entry timing; this study isolates the exit effect, so it may understate the
  stop's full-strategy impact.
- **1m vs 1h evaluation.** Baseline CAGR/Sharpe match, but the geometric MDD
  differs from the report's arithmetic `max_drawdown_percentage`. The research
  contract requires geometric metrics.
- **Fill model.** 5 bps slippage + gap at the 1m open. Real crash liquidity can
  be worse; a gap can exceed the stop.
- **Sample size.** 252 back + 148 forward trades; stop statistics for wide
  stops are thin (2–13 stops).
- **No funding PnL**, same as the strategy report. The strategy is usually on
  the receiving side of funding, so this is conservative.

**Implemented.** The 8% stop is now in the official strategy (1h entries +
5m stop). The updated report shows back Sharpe 1.52 / forward 1.52 and worst
bars of −5.1% / −3.9%, versus −15.9% / −5.0% without the stop.

## 7. Files

- `src/study.py` — full simulation
- `results/summary.csv` — variant table
- `results/trades.parquet` — per-trade time-exit vs stop outcome
- `assets/stop_comparison.png` — forward CAGR / MDD / worst trade by variant
