MTF Research · Phase 0 contract · Review & decide
Lock the MTF research contract.
Phase 0 of the multi-timeframe SMC-primitive investigation. Each task is a methodology decision Claude + GPT converged on. YES = accept the locked default, MODIFY = override (write what), NO = skip / cut from scope. Submitting starts Phase 1 (engine build). No code, no commits, no trading until you submit.
Subject BTC first, Gold validation Phase 7
Decisions 32 tasks · ~5 min
Output reusable engine + first findings
Risk none — no trading, no capital
0 of 32 reviewed
Appendix A — Phase roadmap (after submit)
| Phase | What happens | Output |
|---|
| 1 | Engine foundation (resampling, no-lookahead alignment, primitive labellers, event-study core, bootstrap CI, manifest writer) | scripts/mtf_research/ |
| 2 | Label development data with 4 primitives + state matrix | Labelled dev dataset + state distribution report |
| 3 | Run 7 registered combinations through event study; report pooled + regime-stratified | Event-study CSVs + bootstrap CI summaries |
| 4 | Candidate synthesis from surviving primitives (entry / stop / target as separate modules) | Candidate trade specifications |
| 5 | OOS confirmation — single pass on 2024-2026 data | cohort/mtf_research/ deliverable |
| 6 | Portfolio integration — correlation vs existing book | Deployment recommendation |
| 7 | Gold validation — rerun engine on XAU_USD | Cross-instrument transfer report |
Phases 1-5 are the first execution unit (BTC end-to-end). Phase 6 only triggers if a candidate survives Phase 5. Phase 7 only triggers after Phase 6 has a clear verdict on BTC. Each phase is independently committable.
Appendix B — Methodological lineage
- Three-round Claude ↔ GPT exchange converged on the methodology you're approving here
- Event study before strategy (separate "does the signal exist?" from "does the strategy work?")
- Pre-registered hypothesis space (7 combinations) to control multiple-comparison risk
- 27-cell collapsed state matrix with N≥30 sample-size gating (prevents starved cells from generating noise edge)
- Drift-adjusted shorts (BTC long-only forward return as opportunity-cost baseline, not zero)
- Regime stratification with demote-rule (catches pooled edges that are regime-bound noise)
- Reproducibility manifest (engine commit + data hash + random seed) so OOS rerun a year from now produces identical numbers from identical inputs
- Output framing: reusable engine (durable capability), not one-off strategy (fragile artefact)
Appendix C — What I will NOT do
- Will not touch any trading container, YAML, or capital — this is research only
- Will not iterate on OOS data — single confirmation pass, accept what it tells us
- Will not reactively add deferred primitives (BOS, ChOCh, order block) if first-pass primitives fail; that decision is Phase 9's tree
- Will not tune primitive k-thresholds after seeing event-study results
- Will not promote a strategy from the engine to the live cohort without explicit operator approval after Phase 6 portfolio integration