Modelling Tests Under Pressure: How DCF and LBO Cases Are Actually Scored

At some point in a serious IBD or buy-side adjacent process, somebody hands you a laptop, a case pack and a countdown. The modelling test terrifies candidates more than any interview, and for the wrong reason. They fear the finance. What actually fails people is the clock, because a timed model is a test of process under pressure wearing a test of finance as a costume.

I set and marked these tests for years in M&A at BNP Paribas and in TMT at Nomura. Here is what the marker's sheet actually rewards.

The three formats you will meet

• The timed on-site build. Usually 60-90 minutes, a short case pack, and a request: build a DCF, or a simple LBO, and answer 3-5 questions on the output. Tests speed, hygiene and composure.
• The take-home case. 24-72 hours, a richer pack, sometimes a short deck to accompany the model. Tests judgement and presentation, because with time removed as the excuse, sloppy assumptions become character evidence.
• The case at an assessment centre. A compressed build plus a discussion, folded into a final round alongside interviews and group exercises. The modelling is one input into the day; I have written separately about how superdays and assessment centres are scored.

What the marker is actually scoring

Roughly in this order: does the model balance and flow correctly; are the mechanics right; are the assumptions defensible and clearly flagged; is the output usable by someone who was not in the room; and only then, is the answer clever. A plain model that balances, with hard-coded assumptions in a visible colour and a sensible sensitivity table, beats an ambitious one with a broken balance sheet every single time. Nobody staffs the candidate whose numbers cannot be trusted, however elegant the structure.

The DCF traps that repeat

• Cash flow and discount rate mismatch. Unlevered free cash flow discounted at anything other than WACC, or leverage double-counted. The single most common structural error.
• Terminal value doing all the work. If 85% of your value sits in the terminal period, say so and sanity-check the implied exit multiple. Ignoring it reads as not understanding your own output.
• Growth outrunning the economy forever. A perpetuity growth rate above long-run GDP plus inflation needs a defence you do not have.
• No bridge to equity. Enterprise value left hanging without net debt and the walk to a per-share number. The question was almost certainly what is it worth per share.

The LBO traps that repeat

• Sources and uses that do not tie. The first thing a marker checks, because it is 30 seconds to verify and diagnostic of everything else.
• A debt schedule without a cash sweep. Leverage that never amortises, or sweeps that ignore minimum cash, break the returns story.
• Circularity handled by hope. Interest on average debt creates a loop; either switch on iteration deliberately or use opening balances and say so. An unexplained #REF cascade at minute 80 is fatal.
• Returns with no bridge. State the entry and exit multiple, the leverage, and where the IRR comes from: EBITDA growth, multiple expansion, or paydown. A candidate who can decompose the return understands the trade; one who reports 23% and stops does not.

A protocol for the clock

Read the pack twice before touching the keyboard; the second read is where the trap sentence lives. Sketch the tab structure in 2 minutes: assumptions, model, output. Build the skeleton end to end with placeholder numbers so a working answer exists early, then refine. Timestamp your checkpoints: at halfway you want a balancing model, not a beautiful revenue build. Reserve the final 10 minutes for checks and the written answers, because the questions are where the marks are, and an unanswered question scores zero however good the model behind it.

Sanity checks before you submit: does the balance sheet balance in every year, does implied EV over EBITDA look like a real company, do returns move the right way when you flex leverage, and would the output page make sense to someone who never opened the model. Then stop polishing. Submitted and sound beats perfect and late.

And give the output a page of its own. A marker opens the file at the output tab first: a valuation range rather than a single number, the 2 or 3 sensitivities that actually move it, and the key assumptions stated next to the answer they drive. A model that ends in a range with drivers reads as judgement; a model that ends in one proud number reads as a student exercise, whatever the quality behind it.

Take-homes deserve one extra discipline: a short assumptions note, three lines in the email or on the front tab stating your key calls and why. It converts every debatable assumption from a potential error into a documented judgement, and it is the cheapest way to look like someone who has already worked on live deals.

How to actually prepare

Reps under a timer, not tutorials on a sofa. Build the same 3-statement-to-DCF and simple LBO from a blank sheet 5-10 times until the mechanics are muscle memory and your mind is free for the judgement calls, because judgement is the only thing the test cannot be coached out of. Then rehearse the discussion layer: every model you build should end with you explaining it aloud in 2 minutes, since the conversation about the model is scored as heavily as the file, and it draws on exactly the skills I covered in how to talk about a deal. Sector processes add their own metrics on top, and for technology groups I have set out what TMT interviews actually test.

FAQ

Raphael Tressieres

Former Executive Director in TMT Investment Banking at Nomura and M&A banker at BNP Paribas. Top-rated Head Mentor on Wall Street Oasis with 300+ sessions. About