Signal-only research · No auto-trading · Not financial advice
Optimised Strategy — 20-Day Breakout + Chandelier Trail
The best configuration found by systematic testing: 40-day lookback, 15-day trail, 4×ATR chandelier trailing stop, crypto-perps only. Validated out-of-sample by walk-forward analysis. Backtested on 5 years of real data (2021-08 → 2026-08), net of 21 bps round-trip costs.
1 · How It Works MECHANICS
- Signal: long when a symbol's high breaks its 40-day high; short when its low breaks the 40-day low. The longer lookback (vs the original 20) filters out noise breakouts.
- Ranking: volatility-normalised strength (close − 40d SMA) / ATR(14); enter the top 5 longs + top 5 shorts per day.
- Entry: next bar's open (no look-ahead).
- Exit — chandelier trail: the stop sits 4 × ATR(14) below the highest close since entry (longs) / above the lowest close (shorts). This is the key difference from the original 10-day fixed trail — it adapts to volatility and lets winners run far longer (mean hold 6.4 → 33.5 days) instead of being stopped out on routine pullbacks.
- Costs: 5.5 bps taker + 5 bps slippage per side = 21 bps round-trip, applied to every trade.
Why it wins: the original 10-day trail exits on ordinary pullbacks, so it catches many small wins and a few big ones (17% win rate). The 4×ATR chandelier ignores pullbacks smaller than 4× daily volatility — it only exits on a structural reversal. Result: 41% win rate and the winners run ~3× longer, more than doubling the per-trade edge (LONG avg R +0.49 vs +0.11).
2 · Results vs the Original 5Y · NET · 21 BPS RT
| Config | Trades | Net ret | Win rate | PF | Avg R | Account maxDD |
|---|
| Original 20/10 trail | 1,726 | +2,098% | 17.3% | 1.33 | +0.07 | 19.0% |
| Optimised 40/15 chandelier 4×ATR | 772 | +3,911% | 41.1% | 1.53 | +0.27 | 33.8% |
| TOP3 concentration variant | 1,119 | +2,905% | 16.3% | 1.68 | +0.14 | 11.8% |
Trade-off: +86% more return (+3,911 vs +2,098%) at the cost of ~1.8× the account drawdown (33.8% vs 19.0%) and ~55% fewer trades (772 vs 1,726). At constant 1% risk per trade, the optimised version compounds ×6.48 over 5y (+45.3% CAGR) vs ×2.84 (+23.2%) for the original — and the longer holds mean far less trading effort.
3 · Compounded Account Result 1% RISK / TRADE · NET
Each trade sized as a constant 1% of equity risked, compounded via net R-multiples — what a real low-leverage account following the strategy would have seen. 10,000 block-bootstrap sims.
Actual 5y (1% risk)
+45.3% CAGR
Equity after 5y×6.48
Max DD33.8%
MC distribution
+44.3% CAGR P50
P5 → P95+11.4% → +98.5%
Equity P5 / P50 / P95×1.7 / ×6.3 / ×30.9
P(profit)99.3%
Drawdowns
24.4% P50
MC maxDD P9036.0%
Trades772
At 1% risk/trade, even the 5th-percentile outcome is +11.4% CAGR and 99.3% of sims are profitable. The max drawdown is real (24–36%) — this is a higher-octane configuration than the original, and sizing can be dialled down (0.5% risk ≈ half the DD) if that's preferred.
4 · Long vs Short FULL PERIOD · NET
| Side | Trades | Net ret | Win rate | PF | Avg R |
|---|
| LONG | 405 | +4,137% | 38.3% | 2.07 | +0.49 |
| SHORT | 367 | −226% | 44.1% | 0.94 | +0.02 |
The optimised version makes the long/short asymmetry louder: longs contribute all the P&L (PF 2.07, avg R +0.49) while shorts are a net drag (−226%). A long-only variant (dropping shorts entirely) is the obvious next refinement — expect roughly the long column's +4,137% with ~half the trades and much lower drawdown.
5 · Breakdown by Year NET
| Year | Trades | Net ret | Win rate | PF |
|---|
| 2021* | 22 | +11% | 59.1% | 1.07 |
| 2022 | 116 | −260% | 31.0% | 0.78 |
| 2023 | 138 | +1,810% | 45.7% | 2.85 |
| 2024 | 163 | +688% | 41.7% | 1.39 |
| 2025 | 175 | +1,199% | 41.7% | 1.65 |
| 2026* | 158 | +464% | 40.5% | 1.32 |
(* partial windows of the 2021-08 → 2026-08 data.) Same regime pattern as the original: 2022 (bear) is the only losing year, 2023 is the monster year (+1,810%). The optimised version is profitable in 5 of 6 years and shows a higher win rate in every year than the original's ~17%.
Yearly net return — original vs optimised
Original 20/10
Optimised 40/15 chandelier
6 · Walk-Forward Validation OUT-OF-SAMPLE
Expanding windows: train on earlier years, then measure the config on the unseen following year. If an improvement only shows in-sample, it's curve-fitting; here the chandelier variant beats the baseline out-of-sample in every window.
| Window (test year) | Optimised OOS net | Base 40/15 OOS net | Optimised OOS PF | Base OOS PF | Optimised OOS win |
|---|
| W1 (2023-08 → 2024-08) | +2,119% | +1,515% | 2.95 | 2.80 | 47.2% |
| W2 (2024-08 → 2025-08) | +727% | +237% | 1.45 | 1.24 | 42.1% |
| W3 (2025-08 → 2026-09) | +1,339% | +932% | 1.65 | 1.77 | 41.4% |
Verdict: the chandelier variant beats the base 40/15 on raw return in all 3 OOS windows (+2,119 vs +1,515; +727 vs +237; +1,339 vs +932) and on PF in 2 of 3. The improvement generalizes — it is not in-sample curve-fitting. Caveats: its account DD is higher (7.8–18.3% at 1% risk), and the TOP3 concentration variant's advantage does not hold OOS (roughly neutral), so top-5 stays the recommended setting.
7 · Methodology & Caveats READ ME
- Data: 43 Bybit USDT perps with full 5y history (2021-08-31 → 2026-08-31). Survivorship bias — only currently-trading symbols are included; many alts listed later are absent.
- No look-ahead: signals computed on completed bars; entries at next bar's open; stops use preceding bars only.
- Fills: stop fills at the stop level (adverse-gap fills at open); chandelier fills assume the stop price is reachable intraday.
- Costs: 21 bps round-trip (5.5 fee + 5 slippage per side) — generous for top-tier perps.
- Drawdown metric: account-level, compounding net R-multiples at 1% risk (equity floored at 0 = ruin). Not the unbounded sum-of-returns line.
- Sum-of-returns caveat: the §2 terminal figures are the cross-sectional edge metric and overstate a real portfolio; §3 compounded figures are the realistic account picture.
- Parameter sensitivity: the 4×ATR chandelier is sensitive to the multiplier (3.0 → +2,151%, 2.5 → +698%, 2.0 → +241% on 40/15). It is the best tested configuration, not a guaranteed optimum.
- Signal-only: research. No auto-trading. Not financial advice.