Financial Analysis interview preparation

Questions

100

Traced to a firm

42

Firms

28

Updated

September 2026

Asked at

All firmsMoody's7Bain Capital3SSState Street3AMAres Management2BLBlackRock2DED.E. Shaw2MSMorgan Stanley2Oaktree Capital Management2S&P Global2Bridgewater Associates1Citadel1FTFranklin Templeton1Golub Capital1HWHarris Williams1Houlihan Lokey1J.P. Morgan1Jane Street1MWMarshall Wace1Millennium Management1Morningstar1PIMCO1Sycamore Partners1TSTruist Securities1Two Sigma1Vanguard1WMWellington Management1Wells Fargo Securities1Wolverine Trading1

Topic

All topicsThree statements9Accounting policy and standards5Working capital and cash7Ratio analysis8Forecasting and budgeting9Variance and management reporting7Unit economics and costing8Capital budgeting7Cost of capital and valuation7Markets and rates5Modelling, Excel and data8Business partnering6Brainteasers and estimation4Fit and career10

Level

AnyCoreIntermediateHard

Type

AnyTechnicalCaseBrainteaserMarket viewFit

1. 073What makes a good financial model?Modelling, Excel and dataIntermediatetechnicalFinancial modellingCorporate FP&A

   Say this

   Someone else can pick it up, understand it in ten minutes and change an assumption without breaking it. That means one input layer, consistent formulas across each row, no hard-coded numbers inside calculations, and checks that fail loudly.

   Then walk it

   1. Structure first: inputs, calculations and outputs separated, ideally on separate sheets. One assumption lives in exactly one cell and is referenced everywhere else. If a tax rate appears in three places, the model is already wrong.
   2. Consistency across the row. The same formula copied right across every period, no exceptions, so you can audit a row by checking one cell. A single different formula mid-row is the most common source of silent error.
   3. Formatting as a communication tool: blue for inputs, black for formulas, green for links to other sheets, units labelled on every row, and signs consistent so costs are always negative or always positive, never both.
   4. Checks that are visible. Balance sheet balancing to zero, cash flow tying to the balance sheet cash movement, sources equalling uses, and a single master check cell at the top of every sheet that turns red. A check nobody can see is not a check.
   5. Simplicity. Two hundred rows that a director understands beats two thousand that only you do. If a level of detail does not change a decision, it is cost without benefit, and the honest test is whether removing it moves the answer by more than one percent.
   6. And documentation: a cover sheet with purpose, version, author, date, source of each key assumption and a list of known limitations. In an audit or a diligence process that page saves days.

   Where candidates lose it

   Answering with a list of Excel techniques. The interviewer is asking about design for other people: one input layer, row consistency, visible checks, and restraint on complexity. Mentioning the single master check cell signals you have actually built models in a team.

   Expect next

   • How would you document a model for handover?
   • What check would you build first?
   • When is complexity worth it?
2. 074You inherit a model from someone who left last week and the CFO wants to use it on Monday. How do you audit it?Modelling, Excel and dataHardcase studyFinancial modellingBig Four

   Say this

   Work from the outside in. First sanity-check the outputs against reality, then trace the three or four numbers that drive them, then stress-test rather than read every cell. You cannot audit 40,000 formulas by Monday, so you audit the ones the answer depends on.

   Then walk it

   1. Start with the outputs. Does revenue growth, margin and cash flow look plausible against history and against the sector? An implied 45 percent EBITDA margin on a distribution business tells you more in ten seconds than an hour of cell tracing.
   2. Then the structural checks: does the balance sheet balance, does the cash flow tie to the cash movement, does historical data reconcile to the published accounts. Any break here and nothing downstream is trustworthy.
   3. Then mechanical scanning. Use formula view or a spreadsheet audit tool to find hard-coded constants inside formulas, inconsistent formulas within a row, broken links to external files, circular references and unintended ranges. A row that is consistent for eleven periods and different in the twelfth is exactly what you are hunting.
   4. Then stress tests, which are the fastest way to find logic errors. Set revenue growth to zero, set it to 50 percent, set price to nil. If EBITDA does not move sensibly, or the balance sheet breaks under a stress, you have found a hard-code or a broken link.
   5. Then trace the top three value drivers back to their source, and only those. For a DCF that is usually WACC, terminal growth and the revenue build. Verify each against a document, not against another cell in the same model.
   6. And I would tell the CFO on Friday what I had and had not verified, in writing, with a short list of numbers I do not yet trust. Presenting an unaudited model as clean is the career risk here, not the model itself.

   Where candidates lose it

   Saying you would check every formula. There is not time and it is the wrong method. Outputs first, structural checks, mechanical scan for inconsistencies, then stress tests. And say you would disclose what you could not verify.

   Expect next

   • What is the fastest way to find a hard-coded number?
   • You find an error that changes the answer 20 percent. What do you do?
   • How would you hand this model on properly?
3. 075Your three-statement model does not balance. How do you find the break?Modelling, Excel and dataIntermediatetechnicalFinancial modellingCorporate FP&A

   Say this

   Find the first period where the imbalance appears, then look at the size of the difference, because the number usually names the culprit. Most breaks are one of four things: net income not flowing to retained earnings, a balance sheet movement missing from the cash flow, a sign error, or dividends and capex mishandled.

   Then walk it

   1. Step one, locate. Check the balance row across all periods and find the first column that breaks. Everything after it is contamination; the error is in that one period.
   2. Step two, read the difference. If it equals net income, retained earnings is not picking up the P&L. If it equals twice something, you have a sign error. If it equals depreciation, the add-back is missing or double-counted. The magnitude is the diagnosis.
   3. Step three, check the two mandatory links: net income flows to retained earnings less dividends, and closing cash from the cash flow statement equals the balance sheet cash line. Those two account for most breaks.
   4. Step four, confirm every balance sheet line has a corresponding cash flow movement. A new line added to the balance sheet, a lease liability, a deferred tax balance, an FX reserve, and not wired into the cash flow, is the classic mid-project break.
   5. Step five, check the debt schedule and capex. Gross versus net movements in borrowings, and a capex figure taken from the P&L depreciation rather than the fixed asset schedule, both produce clean-looking models that do not balance.
   6. And the prevention, which is what I would say last: build the balance check from the very first day and keep it on screen. Models that balance from row one never accumulate a break you have to hunt for later.

   Where candidates lose it

   Hunting cell by cell from the top. Locate the first broken period, then let the size of the difference identify the cause. Not knowing that the difference often equals net income or depreciation is what makes this take an afternoon instead of five minutes.

   Expect next

   • The difference equals the depreciation charge. What is wrong?
   • Where would you put the balance check?
   • What if it balances but the cash flow does not tie?
4. 076Your model has a circular reference because interest depends on debt and debt depends on cash flow after interest. How do you handle it?Modelling, Excel and dataHardtechnicalFinancial modellingCorporate finance

   Say this

   Three options: calculate interest on the opening balance instead of the average, use average balances with iterative calculation and a circuit breaker, or resolve it with a small macro. For most work I use opening balances, because the accuracy gain from averages is not worth the fragility.

   Then walk it

   1. The loop: interest expense reduces net income, which reduces cash, which changes the revolver balance, which changes interest. With average debt balances the formula genuinely refers to itself.
   2. Option one, opening balance interest. Clean, no circularity, and typically understates interest slightly in a growing debt year. State the simplification in the assumptions and move on. This is what I would do nine times out of ten.
   3. Option two, enable iterative calculation and build a circuit breaker: a switch cell that forces interest to zero, which lets you break the loop and reset when the model blows up into errors. Without the breaker a single bad input propagates hash-value errors through the whole file and you cannot recover it.
   4. Option three, a macro that copies the calculated interest into the input cell and loops until it converges. Precise, but it means the model needs a macro run to be correct, which is a real handover risk.
   5. The reason it matters beyond aesthetics: circularity makes a model unstable, and an unstable model in a live diligence or board process is a liability. Precision that comes at the cost of reliability is a bad trade.
   6. And the sizing argument I would make out loud: on a 500 crore debt balance at 9 percent, opening versus average balances might move interest by 1 to 2 crore, which is inside the noise of every other assumption in the model.

   Where candidates lose it

   Just switching on iterative calculation with no circuit breaker. That is how models become unrecoverable. Also, pretending average-balance precision matters more than stability shows inexperience; say which trade-off you are making and why.

   Expect next

   • How would you build the circuit breaker?
   • How much does the approximation actually cost you?
   • When would you insist on average balances?
5. 077What is your favourite Excel function?Modelling, Excel and dataCorephone / first roundSSState StreetTax · Boston · 2022

   Say this

   INDEX with MATCH, or XLOOKUP if the version supports it, because it looks up in any direction and does not break when a column is inserted. Give an actual function, then say what you built with it, because that is the part they are testing.

   Then walk it

   1. Why INDEX MATCH over VLOOKUP: VLOOKUP only looks right, it breaks silently when columns move, and it is slower on large ranges. That answer alone tells an interviewer you have maintained someone else's file.
   2. Then say where you used it. Something like: I used INDEX MATCH on two axes to map 400 cost centres to a reporting hierarchy, so a change in the hierarchy needed one table update rather than re-pointing 400 formulas.
   3. Have two more ready, because the real question is the breadth of your toolkit. SUMIFS for driver-based aggregation, which is the workhorse of any actuals-versus-budget pack. And a pivot table over a clean flat table, which is still the fastest way to interrogate a ledger extract.
   4. If they push further, the ones worth naming are power query for repeatable data cleaning, dynamic arrays like FILTER and UNIQUE, and OFFSET or CHOOSE for scenario switching, plus goal seek and data tables for sensitivity.
   5. And one thing to volunteer: I avoid volatile and hard-to-audit constructions. Nested IFs five deep, merged cells, and formulas that span sheets in long chains all make a model unmaintainable even when they work.
   6. The habit I would mention is building the data layer as one flat table with one row per transaction and no merged cells, because every reporting problem downstream gets easier if the source is clean.

   Where candidates lose it

   Naming VLOOKUP with no reason, or answering with a life philosophy instead of a function. Name one, say why it beats the obvious alternative, and attach it to something you actually built. That takes 30 seconds and passes the screen.

   Expect next

   • Why not VLOOKUP?
   • Show me how you would build an actuals-versus-budget pack.
   • What Excel habit do you consider bad practice?

   Reported by candidates at State Street (Tax, Boston, 2022). Source: Wall Street Oasis.

6. 078Tell me about your Excel and Python skills. What do you use each for?Modelling, Excel and dataIntermediatetechnicalSSState StreetGlobal Data · Boston · 2024

   Say this

   Excel for anything a human has to read, review or change, which is most of finance. Python when the data is too big, too messy or too repetitive for Excel, so extracting and cleaning, reconciliations across large files, and anything I have to run every month.

   Then walk it

   1. Be specific about Excel level. Not 'advanced' but what you can do: three-statement models with circularity handled, driver-based revenue builds, power query transformations, pivot-based reporting packs, sensitivity tables and data validation.
   2. Then Python with named libraries and a real use. Pandas for joining and cleaning, openpyxl or xlsxwriter to write the formatted output back to Excel, requests for pulling from an API. The example is what counts: I built a script that reconciled a 400,000-row ledger extract against a bank statement and produced an exceptions file, which replaced two days of manual matching each month.
   3. Say where you draw the line, because that is the judgement they are testing. A model the CFO will change the assumptions in stays in Excel. A monthly data pull with 15 transformation steps goes into Python. Putting business logic in a script nobody else can read is a handover risk, not a win.
   4. Mention SQL if you have it, because in most GCC and shared service roles the data sits in a warehouse and being able to write your own query removes your dependence on a reporting team.
   5. Add the visualisation layer if it is real: Power BI or Tableau, and specifically whether you have built a data model with relationships rather than just charts.
   6. And be honest about the level. Claiming Python and then failing a simple list-comprehension question is far worse than saying you are comfortable with pandas and still learning object-oriented work.

   Where candidates lose it

   Saying 'advanced Excel and basic Python' with no evidence. Name the functions, name the libraries, and attach one before-and-after with a time saving. Overclaiming is fatal because these questions are usually followed by a live test.

   Expect next

   • What would you never do in Python that you would do in Excel?
   • Do you write SQL?
   • Walk me through the logic of your reconciliation script.

   Reported by candidates at State Street (Global Data, Boston, 2024). Source: Wall Street Oasis.

7. 079Tell me about a time you automated a process in a previous role.Modelling, Excel and dataIntermediatetechnicalCitadelCorporate Development · New York · 2025

   Say this

   Pick one process, quantify the before and after in hours and in error rate, and describe what you changed. The structure is: the manual process, why it was painful, what you built, what it saved, and what you would do differently.

   Then walk it

   1. Choose a process with a number attached. A monthly report that took six hours and now takes twenty minutes is a story. 'I made the file more efficient' is not.
   2. Describe the before properly: where the data came from, how many manual steps, what broke. For example, four exports from two systems, copy-pasted into a template, with a reconciliation done by eye.
   3. Then what you built and why that approach. Power query to pull and transform, a single flat data table, pivots for the output, and a check that flags any variance above a threshold. Say why you chose that over a script, because the reasoning is being assessed as much as the result.
   4. Then the quantified result: hours saved per month, errors eliminated, and the more interesting one, what the time was reinvested in. Saving six hours matters because it turned into analysis somebody read.
   5. Then the part most candidates skip: adoption and handover. Did anyone else use it after you left? I would say I documented it on one page and walked two colleagues through it, because an automation that only you can run is a dependency, not an improvement.
   6. And close with a limitation. Something like: it still needed a manual export because we had no database access, and the proper fix was a warehouse connection I could not get approved. That honesty reads far better than a flawless story.

   Where candidates lose it

   Telling a vague efficiency story with no numbers. And do not claim an automation that removed a control. Interviewers in finance care that the checks survived, so say explicitly what validation you built in.

   Expect next

   • How did you make sure the output was right?
   • Did anyone else keep using it?
   • What would you automate next?

   Reported by candidates at Citadel (Corporate Development, New York, 2025). Source: Wall Street Oasis.

8. 080How would you make money with social media data?Modelling, Excel and dataHardcase studyTwo SigmaGeneralist · New York · 2024

   Say this

   Find a place where social data leads a number somebody pays for, and prove the lead is real before you monetise it. The two honest routes are forecasting revenue ahead of official data, and selling the cleaned signal to people who trade or plan on it.

   Then walk it

   1. Start from the decision, not the data. Who pays for being two weeks early? A consumer company deciding on promotional spend, a retailer planning inventory, an investor sizing a quarter. Each of those is a different product.
   2. The forecasting route: mention volume, sentiment and follower growth for a brand, mapped against reported quarterly revenue. If mentions of a product line lead revenue by four weeks with a stable relationship, you have a nowcast that beats consensus.
   3. Then validate it properly, because this is where the question is really marked. Back-test out of sample, check the relationship holds across at least two or three cycles, and control for the obvious confounders: a viral post is not demand, a paid campaign inflates mentions, and bot activity is a large share of raw volume.
   4. Then the monetisation models: sell the cleaned data feed, sell a derived index, or use it internally to trade or to plan. The data business scales better and is easier to defend legally; the internal use captures more of the value but only while the signal stays private.
   5. Then the hard constraints, which a good answer names unprompted. Platform terms of service and API restrictions, privacy law, no personal data, no material non-public information, and signal decay: any edge from public data erodes as more people buy the same feed.
   6. So my proposal would be a narrow pilot. One sector, five brands, six quarters of history, one testable claim: can we predict reported revenue better than consensus. If the back-test fails, stop, because most alternative data claims do not survive it.

   Where candidates lose it

   Jumping to 'sentiment analysis predicts stock prices'. Everyone says it and almost nobody validates it. The marks are in the back-test, the confounders such as bots and paid campaigns, the legal constraints, and naming who pays for the output.

   Expect next

   • How would you back-test that claim?
   • What would you do about bot-generated volume?
   • What happens to the edge once you sell the data?

   Reported by candidates at Two Sigma (Generalist, New York, 2024). Source: Wall Street Oasis.

Firm tags come from public, anonymous candidate reports on Wall Street Oasis: strong signal, not sworn testimony. Firms are named as the places a question was reported, not as partners of Fin Maverick. Answers are written for this page to show how to think out loud; they are not scripts to recite.