project: anomaly-atlas document: expG_cost_frontier/analysis author: Simon-Pierre Boucher contact: contact@spboucher.ai data_source: hfmarketdata.io created: 2026-08-12 modified: 2026-08-12 status: reviewed
Analysis — expG_cost_frontier
Run: results/expG_cost_frontier/20260812T072128Z/results.json (pool rebuilt
mechanically from committed expC/expD/expF outputs — no hand-picking; cost
model and κ sweep pre-declared; cache-served).
The frontier
| κ (× half-spread paid per trade) | survivors / 31 |
|---|---|
| 0 (gross) | 31 |
| 0.1 | 3 (AXDX 30min, CKX 1day, HTD 5min — all sparse names) |
| 0.25 | 1 (CKX 1day, κ* = 0.28) |
| 0.5 | 0 |
| 1.0 | 0 |
Median breakeven κ* = 0.0114: the median double-filtered rule captures ~1 % of one half-spread per trade. The intraday reversion cells sit at κ* 0.004–0.04; the 2014-2015 lead-lag residuals at 0.003–0.03; ES→SPY at 0.0028 (identical across all three splices). Zero intraday rules survive κ = 1 — the pre-registered falsification clause did not trigger.
Reading
- The three-layer doctrine closes. expF showed statistical correction cannot detect mechanism; expG shows the mechanism's price: the gross survivors were harvesting exactly the thing they would have to pay. Charter Q5 answered on this pool: the cost frontier sits an order of magnitude below the most optimistic execution assumptions (Frazzini-style κ ≈ 0.1–0.25).
- The lone κ=0.25 survivor is the skeptic's case study. CKX (ultra sparse; wide, noisy EDGE spread; daily contrarian) has the classic profile of estimation artifact rather than economics. It is NOT discarded by hand — it goes to expH's validation split carrying the skeptical prior, which is what the protocol is for.
Hand-off
expH evaluates on the untouched validation split: (i) CKX 1day (the lone cost survivor), (ii) the daily 2008-2015 reversal family (gross, cost-marginal — evaluated for the decay/negative record), and (iii) the NEGATIVE finding itself ("nothing intraday survives costs") — which, if it replicates out-of-sample, becomes the atlas's first confidence-labeled entries (charter result-types C and E).