workflow-data-pre-modelling-checklist

# Pre-Modelling Data Quality Checklist ## What this skill covers A terminal quality gate that aggregates the outputs of the three upstream data hygiene skills into a single checklist and produces a pass/fail verdict. When this checklist passes, the dataset is cleared for modelling. When it fails, the checklist identifies which dimension requires remediation and routes back to the appropriate upstream skill. ## Inputs - Cleaned dataset with `lineage` and `review_flag` columns populated - Lineage manifest and sign-off block from `workflow-data-lineage-tracking` - Flag summary table from `workflow-data-outlier-detection` - Conversion log and frequency alignment record from `workflow-data-unit-reconciliation` ## Workflow ### Step 1 — Structural profile Confirm the dataset scope matches expectations: - Row count and column count match what was ingested - Target date range is complete (no silent gaps in the time series) - All mandatory columns are present (as defined by the downstream model's input schema) ### Step 2 — Unit and sign check Verify the unit reconciliation output: - [ ] All columns are in the target unit ($M, %, bps — per model convention) - [ ] No column has a residual scale anomaly (validated via `variance_analysis` post-conversion) - [ ] Sign conventions are uniform and match model convention - [ ] Frequency is uniform across all series at the target frequency - [ ] No silent mixed currencies in the output dataset - [ ] Stub periods are labelled and modelling treatment is documented ### Step 3 — Outlier disposition check Verify the outlier detection output: - [ ] All `review_flag` entries have been reviewed by the analyst (no open flags) - [ ] Each flag has a recorded disposition: retain / exclude / winsorise / investigate - [ ] No flagged value has been silently deleted (deletions require explicit approval) - [ ] `factor_model` or `rolling_forecast` residual flags are resolved if raised ### Step 4 — Lineage completeness check Verify the lineage tracking output: - [ ] Every cell in the dataset has a non-empty `lineage` entry - [ ] Every destructive transformation has an `approval` record - [ ] Lineage manifest lists all source extracts with pull timestamps - [ ] Sign-off block is present and `Approved for modelling` is set to "yes" ### Step 5 — Cross-source consistency When the dataset merges multiple sources, verify: - Call `working_capital` to sanity-check DSO/DIO/DPO consistency across cleaned periods - Verify that the same line item from two sources (e.g., FMP revenue vs EDGAR revenue) agrees within a rounding tolerance - Flag any residual discrepancy >0.1% as unresolved (requires analyst review before pass) ### Step 6 — Verdict and routing | All checks pass | Dataset is cleared. Produce the sign-off summary and pass to the downstream model. | |-----------------|-------------------------------------------------------------------------------------| | Unit/sign failure | Route to `workflow-data-unit-reconciliation` for remediation. | | Open outlier flags | Route to `workflow-data-outlier-detection` for analyst review. | | Lineage incomplete | Route to `workflow-data-lineage-tracking` to complete provenance annotation. | | Cross-source mismatch | Escalate to `cfa-chief-analyst` for resolution. | ## Output format 1. **Checklist table** — one row per check, with pass/fail status and remediation action if failed 2. **Verdict block** — PASS or FAIL with list of failing dimensions 3. **Model-ready dataset** — the cleaned dataset with sign-off block, ready for consumption ``` DATA QUALITY SIGN-OFF --------------------- Dataset: [name / description] Date: [ISO date] Source extracts: [list] Unit convention: $M (1dp) | Frequency: Annual | Currency: USD Outstanding flags: 0 Approved for modelling: YES Signed: cfa-chief-analyst ``` ## Quality gates - [ ] All six unit/sign checks pass before proceeding - [ ] Zero open `review_flag` entries in the outlier log - [ ] Lineage sign-off block present and marked "yes" - [ ] Cross-source discrepancies resolved within tolerance - [ ] Verdict explicitly stated as PASS or FAIL — not "mostly clean" ## Related skills - `workflow-data-outlier-detection` — upstream; provides the outlier flag table consumed in Step 3 - `workflow-data-unit-reconciliation` — upstream; provides the conversion log consumed in Step 2 - `workflow-data-lineage-tracking` — upstream; provides the lineage manifest consumed in Step 4