Berg composite and strategy synthesis — composite extreme proxy
Rejected qm-tykd.4 · 2026-08-17
Do Berg extreme-composite aggregations (price-boundary + volume/TRIN + breadth/thrust families) predict turning points beyond unconditional and event baselines, and does a realistic T+1 overlay with costs create a robust tradable strategy?
No. Proxy count>=2 composite (dist>4% & |z|>1.5, vol_z>2, |ret10|>5%) beats unconditional (mean fwd20 1.86% vs 0.94%, hit 64.8% false 35.2% n=358, 4.4% sparse) and event 1-of-3 (0.63%) overall, but fails regime gate: 1994-2007 no edge (0.86% vs 0.87% uncond false 43.2%), 2020+ drives headline (4.97% vs 1.27% hit 88% n=67). F2 volume/TRIN zero signals (NYSE TRIN/breadth free history unavailable), F1 alone trails unconditional, false not reduced vs singletons, extra lag T+2 retains (-1.7bps) but overlay with T+1 lag 10bps underperforms SPY: CAGR 8.44% vs 11.00% delta -2.56pp, gross 9.37% still -1.63pp, Sharpe 0.57 vs 0.65, maxDD -59.9% vs -55.2%. 30k-indicator multiple-testing hazard severe.
Provenance
Equity vs S&P (aligned, rebased to 1.0)
Aligned window, both rebased to 1.0 at start — Aligned window, both rebased to 1.0 at start for fair comparison · S&P dashed, experiment solid. X = Date, Y = Equity (rebased to 1.0).
Signal driver & thresholds (x-aligned)
Signal timeline (sparse triggers)
Triggers when overlay takes over base strategy
Stats
| CAGR | 0.00% | Sharpe | 0.00% |
|---|---|---|---|
| Sortino | 0.00% | maxDD | 0.00% |
Disposition
Rejected — prediction edge does not survive subperiod stability and does not become a tradable overlay; no robust strategy queued. All false composites and negative results retained.
Follow-on work
- Berg tradable strategy promotion from robust composite
- Berg proprietary data proxy feasibility
Errors & data gaps
Native Berg 30k daily series unavailable proprietary — all results are public proxies (SPY/^GSPC + SPY volume + 10d momentum thrust) labelled proxy per ledger; no Berg native performance claim NYSE TRIN/breadth free history incomplete — F2 volume/TRIN proxy yields zero signals under strict vol_z> thresholds, so composite is effectively F1+F3; documented as untestable missing input No walk-forward threshold selection — subperiod splits only; 30k-indicator multiple-testing hazard severe, recent 2020+ outperformance n=67 not stable Overlay is simple cash switch (composite true -> BIL else SPY) not optimized risk-on/off allocation; turnover 8.5/yr trivial but signal sparse 4.4% limits impact
Follow-up
gh issue create --title "Follow-up Berg composite and strategy synthesis composite extreme pro" --body "Follow-up to qm-tykd.4: Do Berg extreme-composite aggregations (price-boundary + volume/TRIN + breadth/t" --label "priority:p2" --label "pipeline:research" --label "agent:ready"