Volatility-Scaled Sizing on the Momentum Strategy — Backtest Report ✅ IMPROVES
Date: 2026-06-12
Analyst: Claude (for Jacques)
Engine: /root/backtest_xsmom_PY-00027_100626.py (unchanged) · Test: /root/xsmom_volscale_120626.py
Builds on: 2026-06-10_cross-sectional-momentum.md (the strategy that PASSED) and
2026-06-12_top20-universe-and-april-whipsaw.md (the April whipsaw diagnosis).
1. Goal
The passing strategy (regime-gated long/short, 21d momentum, top-100 PIT, 50-MA) had one bad month — April 2026, −20.9%. Diagnosis: the long leg equal-weights coins that can move ±90%/week (ARIA, RAVE, SIREN), so a single blow-off dominates the book. This tests the structural fix: inverse-volatility position sizing — bet less on wild coins, more on calm ones, with total gross exposure held equal to the equal-weight baseline (no added leverage, still ≤8 positions). Costs scale with notional. Tune <2026 / OOS 2026 separate.
2. Results — headline vs baseline
| Config | Tune | OOS 2026 | April |
|---|---|---|---|
| Equal-weight (current) | PF 1.21, +45%, DD −31.8%, Sh 0.76 | PF 1.26, +19.8%, DD −20.9%, Sh 1.05 | −20.9% |
| Inverse-vol both legs (lb20, cap2×) | PF 1.41, +134%, DD −16.8%, Sh 1.57 | PF 1.31, +21.3%, DD −19.5%, Sh 1.17 | −13.4% |
| Inverse-vol long-leg only (lb20, cap2×) | PF 1.29, +76%, DD −31.4%, Sh 1.03 | PF 1.30, +21.0%, DD −17.3%, Sh 1.21 | −8.7% |
Full mandatory metrics, recommended config (inverse-vol both legs, vol_lb=20, cap=2×):
| Positions | Win rate | Avg R / pos | PF | Return | Max DD | Last-14-day | |
|---|---|---|---|---|---|---|---|
| Tune (<2026) | 228 | 50.9% | +1.98% | 1.41 | +133.8% | −16.8% | +0.00% (regime off) |
| OOS (2026) | 66 | 59.1% | +2.86% | 1.31 | +21.3% | −19.5% | −4.16% |
OOS monthly (recommended): Jan +11.0% · Feb +0.0% (regime off) · Mar +16.9% · Apr −13.4% · May +8.0% · Jun ~0%.
Robustness across the volatility-lookback window (both legs, cap 2×)
| vol_lb | Tune PF / Sh | OOS PF / ret / Sh | April |
|---|---|---|---|
| 14 | 1.30 / 1.23 | 1.50 / +36.4% / 1.74 | −10.4% |
| 20 | 1.41 / 1.57 | 1.31 / +21.3% / 1.17 | −13.4% |
| 30 | 1.36 / 1.34 | 1.34 / +23.5% / 1.23 | −13.5% |
| 45 | 1.38 / 1.37 | 1.32 / +22.4% / 1.24 | −11.6% |
Every window beats the equal-weight baseline on both tune and OOS — a plateau, not a spike. The 14-day corner looks best OOS but is the overfit-prone edge; the honest center is 20–30 days. The 2× cap barely changes results vs no-cap (the inverse-vol weights weren't over-concentrating into one coin anyway) — it's a cheap safety belt, kept.
3. What it bought us
- April whipsaw cut by a third (−20.9% → −13.4% both legs; −8.7% if only the long leg is scaled). The wild coins now get a sliver of a bet instead of a full one — exactly the intended mechanism.
- Tune drawdown roughly halved (−31.8% → −16.8%). This is the clearest, most credible signal: the risk control does what it's supposed to.
- OOS Sharpe and PF both improved (1.05→1.17, 1.26→1.31), return held (+21% vs +20%).
4. Honest caveats
- The spectacular numbers are in the tune period (+134%, Sharpe 1.57). The out-of-sample gain is modest — the honest headline is risk reduction, not a return explosion. OOS return is essentially unchanged; what improved is drawdown and consistency.
- April is reduced, not cured (−13%). These coins move so violently that even a small weight bleeds. Vol-sizing dampens the whipsaw; it does not erase it.
- OOS is still statistically thin (66 position-weeks, 11 active bull weeks). Confidence comes from tune+OOS agreement and the parameter plateau, not 2026 alone.
- Currently in a small drawdown: last-14-day −4.2% (same as the base strategy — regime recently switched off).
5. Conclusion & recommendation
Inverse-volatility sizing is the first modification this round that improves the strategy out-of-sample while halving tune drawdown and cutting the April whipsaw by a third. It is robust across the vol-lookback window and adds no leverage. Recommended config: regime_ls, top-100, 50-MA, 21d momentum, inverse-vol both legs, vol_lb 20–30, cap 2×.
Open choice: scaling only the long leg gives the best April (−8.7%) and lowest OOS drawdown (−17.3%) but leaves tune drawdown high (shorts still equal-weighted); scaling both legs gives the best tune drawdown (−16.8%). Both-legs is the more principled default.
This strengthens — but does not replace — the forward paper-trade recommendation from the original report. Next candidate threads: (a) a "skip-the-parabolic" long filter to attack the residual April bleed; (b) confirm inverse-vol holds in a fresh forward week.
Status: PASS — improves the validated strategy. Adopt inverse-vol sizing for the forward test.