Trading Book · crypto research / 28 September 2026
Crypto thresholds:
what survived the six-month replay
1,536 seeded threshold profiles, exact hourly trade and mark execution, $100 risk per trade, and a locked-period failure that changes the conclusion.
The answer
There is no honestly validated optimum yet. The automatically selected development winner lost $-377.29 in the locked August–September period. The balanced profile below is the best paper candidate after inspecting the complete six months; it is not ready for live capital.
The balanced candidate produced 20 closed trades, nine wins (45%), $+1,315.69 closed profit, $+196.68 open P&L after reserved exit costs, profit factor 3.37, and 7.4% maximum hourly-close drawdown. It preserves the requested high reward/risk: a six-historical-ADR target and at least 4R at plan and simulated fill.
The crypto numbers
| Gate | Balanced crypto threshold | Logic retained |
|---|---|---|
| G1 | Close ≤2% above SMA10 | Do not chase the short moving average |
| G2 | Close ≤8% above SMA20; forgive through 12% when momentum ≥30% | Cap extension but allow exceptional strength |
| G3 | 30-day return ≥0% | Mandatory trend direction |
| G4 | Daily base and contraction ≤8% | Mandatory, objectively bounded setup |
| G5 | 30-day return minus BTC ≥−10 points | Crypto leadership context |
| G6 | Last completed funding observation ≥−0.01% | Historical setup-time context; missing blocks |
| G7 | 30-day universe percentile ≥60 | Cross-sectional leadership proxy |
| G8 | Historical ADR20 between 1% and 8% | Volatility at entry, never today's ADR written backward |
| Admission | G1–G4 mandatory; at least 6/8 | Current equity tolerance shape; missing inputs never pass |
| Exit | 6 ADR target; minimum 4R | High reward/risk requested by the user |
Every ADR reading uses only the twenty completed daily UTC candles available at that historical signal. BTC-relative strength, return percentile and funding are likewise frozen at the setup timestamp.
Full six-month results and stress
All results use 5bps fees and 5bps adverse slippage on each side plus 3bps/day adverse funding unless the row says doubled costs. The $1,200 notional cap and integer contracts often reduce actual structural risk below $100.
| Profile / scenario | Closed | Wins | Win rate | Closed net | Open net | Total | PF | Drawdown |
|---|---|---|---|---|---|---|---|---|
| 156 / base | 20 | 9 | 45.0% | $+1,315.69 | $+196.68 | $+1,512.36 | 3.37 | 7.4% |
| 156 / double_costs | 20 | 9 | 45.0% | $+1,234.56 | $+187.84 | $+1,422.41 | 3.10 | 8.0% |
| 156 / one_hour_later | 23 | 10 | 43.5% | $+1,223.92 | $+28.01 | $+1,251.93 | 2.84 | 7.7% |
| 156 / risk_to_disaster | 20 | 9 | 45.0% | $+1,315.69 | $+196.68 | $+1,512.36 | 3.37 | 7.4% |
| 226 / base | 24 | 8 | 33.3% | $+1,546.29 | $+196.68 | $+1,742.97 | 2.78 | 8.1% |
| 226 / double_costs | 24 | 8 | 33.3% | $+1,456.20 | $+187.84 | $+1,644.05 | 2.59 | 8.3% |
| 226 / one_hour_later | 26 | 8 | 30.8% | $+1,465.89 | $+28.01 | $+1,493.90 | 2.54 | 8.1% |
| 226 / risk_to_disaster | 24 | 8 | 33.3% | $+1,546.29 | $+196.68 | $+1,742.97 | 2.78 | 8.1% |
The balanced candidate retained $+1,234.56 closed profit under doubled costs and $+1,223.92 with entry delayed one hour.
Why this is still paper-only
The first 768 profiles copied equity-scale ranges. None qualified. The June–July development section had only five BTC-above-EMA50 dates, so treating each calendar half as if it should trade equally was wrong.
The expanded search tested another 384 runners and 384 high-R profiles. Exact-hourly development ended on 31 July. A profile needed at least 15 closed trades, positive closed profit, profit factor at least 1.2, drawdown no more than 25%, and a positive one-sided 90% lower confidence score. No equity-style uncapped runner qualified. Eight high-R profiles did.
The locked winner had 20 development trades, six wins, +$623.83 closed profit and 8.0% drawdown. In August–September it produced six closed trades, one win and $-377.29 closed profit. Two open positions marked at $+196.68, leaving total P&L $-180.61. That is a failed locked test.
Seven other qualifying profiles were positive in the locked-period diagnostic, but selecting among them after seeing those outcomes would be data snooping. The balanced candidate is therefore a frozen forward-paper hypothesis. Changing its thresholds during forward observation restarts validation.
All eight development-qualified profiles over the complete, already-inspected six months
| Profile | Closed | Wins | Win rate | Closed net | Total | PF | Drawdown |
|---|---|---|---|---|---|---|---|
| expanded_high_r_009 | 22 | 7 | 31.8% | $+1,129.23 | $+1,325.91 | 2.37 | 9.8% |
| expanded_high_r_026 | 21 | 7 | 33.3% | $+1,225.81 | $+1,422.49 | 2.68 | 8.4% |
| expanded_high_r_101 | 18 | 6 | 33.3% | $+1,251.92 | $+1,448.60 | 2.88 | 8.3% |
| expanded_high_r_122 | 21 | 6 | 28.6% | $+1,158.57 | $+1,355.24 | 2.47 | 8.0% |
| expanded_high_r_156 | 20 | 9 | 45.0% | $+1,315.69 | $+1,512.36 | 3.37 | 7.4% |
| expanded_high_r_183 | 27 | 8 | 29.6% | $+298.82 | $+495.49 | 1.28 | 10.0% |
| expanded_high_r_226 | 24 | 8 | 33.3% | $+1,546.29 | $+1,742.97 | 2.78 | 8.1% |
| expanded_high_r_291 | 21 | 6 | 28.6% | $+649.25 | $+845.92 | 1.70 | 7.4% |
Reproduction and evidence
The offline command now accepts crypto-specific SMA extension, momentum, BTC-relative, funding, percentile, ADR, gate-count, mandatory-gate and reward/risk parameters. The balanced command reproduced the saved summary exactly: 20 closed, nine wins and $+1,512.36 total P&L.
Verification: 808 tests passed with one existing skip. Changed source files pass Ruff and strict mypy. The repository-wide Ruff check is blocked by a pre-existing long line in crypto_service/tasks.py; repository-wide mypy is blocked by two pre-existing errors in backtest_tis_chop.py.
Downloads: summary · full finalist table · locked frontier diagnostic · cost and latency stress · selection lock · command replay.
Production configuration was not changed, no order was placed, and both agent processes remain intended to stay at zero dynos.