Kumu / Corporate Finance / Stage 5
Stage 5 — Analysis, Evaluation & Repo Polish
The capstone. Feed your spec — and nothing else — to an LLM. Verify its numbers by hand. Correct, annotate, and finish with the executive judgment only a human can supply. The future of finance work isn't "do the analysis" or "let AI do it" — it's specify, evaluate, and take responsibility.
The three moves
- Execute. Paste your Stage 4 spec into a fresh LLM session with zero extra context — no "please also consider…". Save the complete raw output. This is your spec's exam.
- Verify & evaluate. Recompute ≥5 ratios by hand from your Stage 3 financials, compare to the LLM's values, and flag every discrepancy with a cause. Then edit the analysis: correct errors, annotate weak reasoning, add your voice.
- Polish. Make the repo a portfolio artifact — READMEs everywhere, clean history, license, description. Budget 2–3 hours; don't leave it for the night before.
The verification table — your defense against confident nonsense
Three sources of ratio values exist by now: the template's auto-computed values, the LLM's stated values, and your manual recomputation. The graded table compares manual vs LLM. Pick ratios across categories, and deliberately choose the ones an LLM is most likely to fumble: anything involving averages, start-of-year values, or unit conversions.
| Ratio | Formula (named ranges) | Manual (show arithmetic) | LLM | Match? | Note |
|---|---|---|---|---|---|
| ROA | INC_net_income / avg(BAL_assets_total) | 12,500 / ((100,000+96,000)/2) = 12.76% | 12.76% | ✓ | Correct averaging |
| DSO | (BAL_receivables / INC_revenue) × 365 | (8,000/80,000) × 365 = 36.5 d | 36.0 d | ✗ | LLM used a 360-day year without disclosing |
| Inventory turnover | INC_cogs / avg(BAL_inventory) | 60,000 / ((15,000+13,000)/2) = 4.29× | 4.62× | ✗ | LLM used end-of-year inventory, not the average |
Discrepancies are the good rows
A mismatch with a one-line causal note earns full credit — it proves you caught the machine. An unflagged discrepancy is the failure mode. (And every mismatch is retrospective gold: did your spec state the day-count convention? The averaging rule?)
The six deliverables
| # | File | Where |
|---|---|---|
| 1 | Raw LLM output (unedited) | deliverables/…-llm-raw.md |
| 2 | Manual verification table (≥5 ratios) | analysis/validation/…-stage5-verification.md |
| 3 | Evaluated final analysis | deliverables/…-final-analysis.md |
| 4 | Spec retrospective (template-based) | deliverables/…-spec-retrospective.md |
| 5 | Updated prompt log | deliverables/prompt-log.md |
| 6 | (Optional) Stage 2 feedback response | docs/decisions/…-stage2-feedback-response.md |
Produce them in order 1 → 2 → 3 → 4 → 5: raw output first, verification second — so you confront discrepancies before writing the final analysis, not after.
The final analysis must contain
- Company & data summary — verified context, assumptions, accounting-standard notes
- Ratio results & interpretation — all six categories, with your corrections to the LLM
- Du Pont analysis — is the LLM's decomposition story sound?
- Strategic recommendations — 3–5, each citing specific ratio values; note what nuance the LLM missed
- LLM evaluation — what it executed correctly; where it deviated, hallucinated, or oversimplified; spec-gap errors vs model-limitation errors
- Executive justification — the investment/strategic thesis in your voice. The "so what?" no machine supplies.
The spec retrospective — honest beats flattering
Using the course template: a section-by-section verdict (Clear / Vague / Missing) tied to symptoms in the output, your top three gaps with evidence, three revisions you'd make, an effectiveness rating, and one process suggestion. "My spec was perfect" scores worse than "Part B.9 said 'recommend strategic actions' without an evidence standard, so the LLM produced generic advice — I'd require each recommendation to cite two ratio values."
Feedback incorporation & repo polish
- Stage 2 feedback (5%): show a visible response to the instructor's PR comments — commits that reference them, a revised memo, or a follow-up memo. Disagreeing is fine if the disagreement is visible; silence reads as oversight.
- Polish (15%): top-level README with project status and commit hashes, a
one-line repo description in the GitHub header, a LICENSE (MIT/Apache-2.0),
.gitignore, README in every directory, filenames on convention, no orphan files, clean commit history, and — still — public.
Submission checklist
- Spec executed cold — raw output saved unedited to
deliverables/ - Verification table: ≥5 ratios across categories, arithmetic shown, discrepancies explained
- Final analysis complete with all six sections, corrections annotated
- Spec retrospective done on the template — gaps tied to evidence
- Prompt log updated with the Stage 5 sessions
- Stage 2 PR feedback visibly incorporated (or visibly, reasonably rebutted)
- All earlier-stage artifacts consolidated in the repo (no Lamaku-only stragglers)
- Polish pass: README status + hashes, repo description, LICENSE, per-dir READMEs, clean history
- Repo URL submitted on Lamaku
Rubric
| Criterion | % | Strong work looks like |
|---|---|---|
| Analytical correctness | 25% | Numbers tie to source; interpretations defensible; LLM errors caught and fixed |
| Manual verification artifact | 10% | ≥5 hand-recomputed ratios; discrepancies flagged and explained |
| LLM evaluation + spec retrospective | 25% | Specific, evidence-tied critique of the AI and of your own spec |
| Strategic recommendations + executive voice | 20% | Each rec backed by ratio evidence; the voice is yours |
| Stage 2 feedback incorporation | 5% | Visible response to PR comments |
| Repo polish | 15% | A professional artifact you'd link on LinkedIn |