Composites comparison
Every composite variant we've built, head-to-head. Same panel, same costs.
Source:
scripts/best_of_combo.py, composite_industry_relative.py, composite_sector_industry_combined.py.
Full window (post-shock + pre-shock)
| Composite variant | Sharpe | Sortino | Calmar | CAGR | MaxDD | Vol | Skew |
|---|---|---|---|---|---|---|---|
Reading guide
- Composite_5sig (flat) = baseline: 5 verified-real-alpha features, pooled IR weights, sector-neutral z-score.
- Per-sector composite = each sector uses its own IR-eff weights for the 5 features. Sign-aware (range_true is negative in tech, positive in materials etc.). +0.49 Sharpe over flat.
- Industry-relative composite = each feature uses its winning variant (abs / vs_industry_median / industry_z / industry_pct_rank). Slightly below per-sector โ 110-industry buckets are too granular.
- Sector ร industry combined = per-sector weights ร industry-relative variants. ~1.95 Sharpe โ comparable to per-sector alone.
- Vol-target overlay (ฯ*=15%) = on top of any composite, scale leverage daily by 60-day realised vol. Cuts MaxDD by 4ร while keeping Sharpe ~constant.
- Annual rebalance = quarterly was over-trading the 252-day-horizon signals. Annual gets +0.16 Sharpe and 60% lower turnover cost.
Pre-2026-05-04 results all use the 2021-2026 panel. Once the 2018-extended panel finishes the multi-cycle regime test will give the definitive verdict on whether per-sector weights are real or in-sample-fit.