Research program map and measured verdict after segment 5 and four parallel tracks, dated 2026-09-27; synthetic success did not establish a real-market edge.
Real-data headline
No positive realistic-cost edge was established on the free real-data fixtures. The synthetic run-#16 book transferred cold at median Sharpe −1.13 flat / −4.69 realistic across BTC 2021–2025 and ETH/SOL 2024. The run-#20 real-data book reached +0.63 TUNE / −0.28 VALIDATE under realistic costs. Eleven catalog members are silent and MVRV is degenerate on the missing on-chain inputs: the synthetic whale/flow claim remains untestable without suitable paid historical feeds, not validated or disproved by those zero-filled columns.
The calibrated simulator agrees in direction: the restored run-#16 book on synthetic family 4 scores +1.64 flat / −0.31 realistic. Cross-sectional ranks also fail after costs. Project-card ranking yields a weak hypothesis, not a proven trading strategy. Read Real-data verdict before interpreting the older synthetic pages.
Program map
| Track | Role and measured result | Read next | Main limitation |
|---|---|---|---|
| Segment 5: real-market time series | BTC 2021–2023 TUNE; BTC 2024/2025 + ETH/SOL 2024 VALIDATE; realistic validation median −0.28 | Real-data verdict and coverage | Free inputs cannot test the synthetic on-chain sleeves; all displayed validation paths are now observed |
| B: simulator calibration | Stylized-fact distance to real falls from f1 4.86 → f2 0.79 → f3 0.55 → f4 0.42 | Program tracks and reproduction | Matched facts are not proof of causal market structure; calibration saw all seven real paths |
| C: cross-sectional long-short | Six daily rank signals on requested 20-perp pool; realistic median TUNE −3.31 / VALIDATE −1.18 | Cross-sectional results | Ex-post pool and retrospective inclusion; costs dominate |
| D: project cards and PIT ranking | 91,807 rows; 2,649 top-500 disappearance labels; validation rank IC +0.061, Newey–West t 1.9 | Project-card evidence | Age is the strongest consistent feature; SENSE/team/audit judgement fields remain unfilled |
| E: raw data lake | Snapshot: 7,525 files / 0.74 GB; free market, TVL, GitHub and CMC quarterly data | Lake coverage and commands | CMC 2,500-entry cap starts at 2020Q3; public access is not a redistribution license |
Pages
- Real-data verdict — segment-5 tables, data coverage, costs, and what paid data would unlock
- Program tracks — lake, cards, calibration, cross-sectional findings and reproduction commands
- Synthetic verdict and decision table — historical segment-4 result and 27-signal weights, not the current real-data verdict
- Regime map and layering — synthetic alpha concentration and stack construction
- Validation and rejections — historical seed roles, run ledgers and rejected candidates
- Harness and costs — historical cost sweeps and real-feed ingest design
Historical synthetic baseline
These are the segment-4/run-#16 measurements previously headlining this page. They describe synthetic generators, not investable market performance; the linked historical pages retain that context.
| Synthetic family | Generator character | Historical status | Flat median Sharpe |
|---|---|---|---|
| 1 | Whale leads always strike | IN1/HO1 burned; fresh single touches 1111001-5 median 3.39, 1212001-5 median 3.79 | 4.817 |
| 2 | Student-t tails, about 30% structurally unidentifiable false whale leads, informative funding after the segment-3 fix | IN2/HO2 burned; fresh single touches 2222001-5 median 1.79/worst −1.21, 2323001-5 median 0.08 | 2.085 |
| 3 | Jump diffusion, latent market beta, separable false lead-ins | Segment 4 closed; run #16 = run #18 book; run #17 descent flagged and reverted | 4.353 |
| 4: calibrated synthetic | Family-3 structure calibrated to real stylized facts | Run-#16 book restored explicitly; realistic median −0.308 | 1.644 |
The runner’s labels need care: fixtures4 is calibrated synthetic data scored with --family 3; fixtures_real is labeled family4 by the runner. These are different datasets, never pooled evidence.
Sources and verification boundary
| Claim set | Source available in the research workspace | Evidence date | Caveat |
|---|---|---|---|
| Segment-5 results and coverage | auto/market-signals/notes/03-validation-ledger.md, fixtures_real/COVERAGE.md; /tmp/ms5-harvest-receipts.md campaign receipt | 2026-09-27 | Local research artifacts, not downloadable from this wiki |
| Four parallel tracks | auto/lake/README.md; auto/project-cards/notes/01-pit-backtest.md; auto/market-signals/notes/05-calibration.md; auto/market-signals/xs/README.md | 2026-09-27 | Snapshot numbers, not live counters; exact caveats and commands on the tracks page |
| Historical synthetic pages | auto/market-signals/CHEATSHEET.md, notes/01-signal-atlas.md through notes/04-stack-map.md, harness/README.md | Segment 4: 2026-09-27, historical book 3108547 | Superseded as the program headline by the real-data null result |
- Workspace root:
~/dev/research/; market-signals/lake/project-cards source snapshot references at consolidation:e5e8dd6/a62e2b6/1f8183a. - The tables report recorded research measurements, not a new rerun and not investment advice. Reproduction commands and split definitions are on the two current pages.
- Local explainer, not hosted here:
/Users/mbp/dev/explainers/2026-09-27_market-signals/index.html.