MeanyDeany · Methodology
Research methodology
Time-respecting construction, narrow model roles, robust comparison, immutable evidence, and visible failure states.
- Data
- Time-respecting
- Models
- Compared, not canonized
- Evidence
- Auditable and immutable
- Permission
- Outside model output
Working principles
Research credibility is cumulative
No single diagnostic establishes a system's validity. Credibility accumulates through temporal discipline, robustness, provenance, operational integrity, and appropriately narrow claims.
- 01Evidence before permission
A model output earns interpretation through validation. It does not become an entry, veto, sizing, or execution instruction by default.
- 02Time-respecting data construction
Features, states, and outcomes use information available at the evaluated timestamp. Alignment rules are explicit and missing data is not silently repaired.
- 03Model comparison and robustness
Ablation, alternative specifications, walk-forward evaluation, and subperiod checks matter more than a single favorable summary statistic.
- 04Immutable evidence and provenance
Historical research records should remain inspectable as hypotheses evolve. Source identity and transformation history belong beside the result.
- 05Operational reliability
Duplicate protection, stale-input detection, locking, scheduling, and failure states are part of research validity when systems generate evidence repeatedly.
- 06Evidence, policy, and execution stay separate
Descriptive evidence can inform later policy research, but policy state is not entry permission and neither layer authorizes execution.
Research notes
Short rules for difficult decisions
- Timestamp discipline
- If a value was not knowable at the decision timestamp, it does not belong in the feature set.
- Narrow claims
- Model maturity, forecast loss, and regime labels answer research questions—not whether a strategy is approved.
- Visible failure
- A failed input, orphan reference, or stale process should remain observable instead of being hidden by a plausible fallback.
- Reproducible lineage
- A result is stronger when another reviewer can trace its inputs, specification, timing, and limitations.
Separation of concerns
One pipeline, three distinct responsibilities
- Evidence
- What the data and model record support.
- Interpretation
- How the evidence is framed and challenged.
- Policy
- A separate research layer with explicit rules.
- Execution
- Not part of this public portfolio.