Vol-target sizing for donchian@1d — path-accurate, full-wallet basis

2026-06-16 · status: inconclusive (vol-target neutral on BTC; lever is exposure level)

Hypothesis: On donchian@1d at the CURRENT full-wallet × lev3 basis (raw maxDD ~50% on BTC), vol-target sizing materially cuts drawdown without losing risk-adjusted return — unlike the 2026-05 re-pricer runs (old ~12%-capped basis, baseline DD ~0.4%, linear PnL approximation).

Verdict: Vol-target adds nothing robust on single-asset BTC. The real lever is the exposure level. Vol-target's value is cross-asset (out of scope).

Why a new run

The donchian cross-asset gate (2026-06-16) cleared donchian@1d (PF 1.30/1.37/1.39 on BTC/ETH/SOL, positive neighborhood plateau on both non-BTC assets) — the first Donchian to pass. But raw drawdowns at full-wallet × lev3 are brutal: BTC 50%, ETH 68%, SOL 92%. Sizing is the whole game.

Prior vol-target experiments (vol_target_sizing, vol_target_strategies) ran on the old ~12%-capped sizing era (baseline DD ~0.4% — nothing to compress) and used a linear re-pricer (sum of scaled per-trade PnL) that cannot model the drawdown path. This run fixes both.

Method (exact for a sizing overlay)

Vol-targeting changes only position size, not signals or stop prices, so each trade's net return-on-notional r_i = pnl_usd / (size·entry_price) is size-invariant. We walk the balance trade-by-trade with real compounding:

notional_i = balance · 0.90 · mult_i           mult_i = clip(target / trailing_RV_i, lo, hi)
balance   += notional_i · r_i                   baseline = same walk, mult_i ≡ 1

trailing_RV = 30-day realised vol (annualised), causal — no model, no leakage.

Sanity gate passed: the mult≡1.0 re-sim reproduces the runner exactly — +323.6% / maxDD 50.3% on BTC. The path model is correct.

Result 1 — vol-target never beats baseline Sharpe

Every vol-target Sharpe (1.04–1.10) sits at or below the baseline 1.16. The realised-vol signal adds no risk-adjusted value — the ATR stop already vol-adapts the risk per trade.

Result 2 — flat deleverage dominates on the robust metric

Pure scaling holds Sharpe flat at 1.16 at every exposure level. So flat deleverage matches any DD target the vol-target rule hits, at equal-or-better Sharpe.

Vol-target does edge flat on Calmar at matched maxDD (7.59 vs ~6.3) — it concentrates its size cuts on the single worst-drawdown window. But that advantage rests on one event at n=49 and is too fragile to bank without a Monte-Carlo reorder confirming it survives bad-luck sequencing.

Takeaways

1. For single-asset BTC: don't vol-target — pick an exposure level. The exposure→DD table above is the decision tool. Target maxDD ⇒ pick alloc. E.g. ~20% DD ⇒ alloc ≈ 0.35 (≈ +15%/yr); ~30% DD ⇒ alloc ≈ 0.55 (≈ +21%/yr).
2. Full-wallet × lev3 is reckless for this edge (50% DD on the friendliest asset). The deployed DONCHIAN_20 only tames this via partial_tp + the ATR trail.
3. Vol-target earns its keep cross-asset, not within BTC: the same alloc gives 50/68/92% DD on BTC/ETH/SOL — there a target-vol rule auto-shrinks the high-vol assets. If donchian@1d ever goes multi-asset, revisit vol-target as the cross-asset normaliser, not as a BTC timing signal.

Caveats

• Single path, n=49 trades, BTC only. maxDD is one event — Sharpe is the trustworthy metric here, and it says vol-target ≈ neutral.
• 30-day realised vol as the proxy (causal). The GBM/HAR forecast (predict_vol_1d) might time vol better, but Result 1's Sharpe ceiling suggests limited upside.
• Returns are 6-year totals (2020–2025); ~27%/yr at alloc 0.90 down to ~11%/yr at 0.25.