MGMT 648 · Session 1
Thu Aug 27, 2026
Capital budgeting
Free cash flow, NPV, IRR, incremental cash flows, multiple alternatives, real options, and simulation. Sept 12
Company valuation
Two-stage DCF model (Sept 17) and multiples of comparables (Oct 1)
Rates and portfolios
Mean-variance optimization, the cost of capital, bond pricing and duration, immunization. Oct 10
Wrap-Up
Live in-class Q&A. Oct 15
Excel
Still the right tool for a tabular model that adds and subtracts, and the one your colleagues can open, change, and argue with.
AI
Everything else: fetching filings, estimating a beta, optimizing a portfolio, running ten thousand simulations. It also builds the workbook.
Everything is posted at mgmt648.kerryback.com.
Schedule
The home page. Slides and files distributed here.
Syllabus
Grading, meeting times, and what to prepare. Nothing outside class is collected.
Free readings
Welch’s Corporate Finance, Damodaran’s materials
Write =B5*(1+$B$6), not =B5*1.05. Source format: document, date, specific reference — “Company 10-K, FY2025, page 45, revenue note.”
| Convention | Rule |
|---|---|
| Blue text | Hardcoded inputs — the cells a reader is meant to change |
| Black text | Every formula and calculation |
| Green text | Links to another sheet in the same workbook |
| Currency | Dollar sign, commas, no decimals: $1,234,567 |
| Units | Stated in the header, not in every cell: Revenue ($mm) |
| Percentages | One decimal: 12.4%. Valuation multiples as 8.5x |
| Negatives | In parentheses: (123), not −123 |
Zero formula errors on delivery: no #REF!, #DIV/0!, #VALUE!, or #NAME?.
The color convention is not decoration. It is how a reader knows, at a glance, what they are allowed to change.
For your reader
Blue says “this is an assumption, argue with it.” Black says “this follows from the assumptions, do not touch it.”
S0...A Linux container on a cloud server, billed to the school. Security on it is minimal — keep nothing personal there.
The harness
Claude Code, connected to GLM 4.7 — a capable and relatively inexpensive model from Z.ai.
The software
Python and its scientific libraries, plus skills for financial data, spreadsheets, and slides.
The screen
File browser and terminal on the left; the prompt window at the foot of the right-hand panel.
Refresh
The file browser does not notice a new file on its own. Use the refresh button above it.
No Excel in the lab
Office is not installed. Double-clicking a workbook there does nothing.
Download to open
Workbooks come down to your laptop through the three-dots menu, and open in your own Excel.
The model cannot open a file, run a calculation, or fetch a filing. It reads text and writes text. That is the entire capability.
What makes it useful is the harness around it. The model emits a request to use a tool; the harness runs the tool and reports the result back.
On the lab, GLM writes a Python file, runs it at the terminal, reads the output, and revises. Four tool calls before you see anything.
Excel workbooks
Written cell by cell with Python: inputs as numbers, formulas as formulas.
Charts and figures
Any chart style, custom annotations, labels, fonts
Word and PowerPoint
Real files with editable text, not pictures of them.
Data analysis
Data fetched, merged, filtered, aggregated — statistics and regressions
Advanced data analysis
Simulations, optimizations, machine learning
Web apps
A page someone else can use, from a description of what it should do.
Try this prompt:
Build an Excel workbook containing a mortgage amortization table. Put the loan amount, rate, and term in labeled input cells at the top. Include a chart of the amount going to interest and the amount going to principal each month.

The workbook
Inputs in labeled cells, an outputs block below them, 360 rows of schedule.
The selected payment cell holds =$B$12, not 2,528.27.
Change the rate and the whole sheet recalculates.
A chart object bound to the cells, not a picture pasted in. Change the rate and the crossover moves with it.
There are also Claude, ChatGPT, and CoPilot plugins for Excel , from Home → Add-ins. The chat window sits inside the workbook.
They need paid accounts or particular types of license agreements. We will not use them.
For any complex operation, ask AI to plan first, or ask it what the best way is to do something, before it begins to work.
Planning pulls the knowledge in
Plan how you will do this before you start. The plan puts the relevant considerations into the conversation, and it answers in light of those rather than from base rates.
Planning breaks the work up
Steps you can check one at a time. Failures stay small and local instead of surfacing at the end, when everything downstream is already wrong.
The problem
Errors usually come from
What to say:
Training on human feedback teaches that agreement earns high ratings. Ask “is this valuation sound?” and it will find reasons that it is.
Invert it. What is the strongest argument that this project should be rejected?
AI is not a good self-critic
Ask AI in the same conversation to evaluate its work. It will review its own reasoning and conclude it is sound.
Start a new conversation or ask for a subagent
A new conversation or a subagent will not see the prior reasoning. It will look at it “with a fresh set of eyes.”
Check the outputs
We don’t usually check AI code by reading it. We check small cases where we can verify “by hand” (using AI to help with that too).
And keep what works
Keep code that has been vetted. Tell AI to use it again the next time.
MGMT 648 · Applied Finance