Portfolio Management interview preparation

Questions

100

Traced to a firm

40

Firms

24

Updated

September 2026

Asked at

All firmsBLBlackRock4Vanguard4WMWellington Management4Amundi3ACAQR Capital Management3Neuberger Berman3SCSchroders3Man Group2MSCI2Northern Trust2AllianceBernstein1Apollo Global Management1Blackstone1BMBNY Mellon1Carlyle Group1Fidelity Investments1Goldman Sachs1Invesco1Millennium Management1MSMorgan Stanley1NUNuveen1PIMCO1SSState Street1TPTPG1

Topic

All topicsPortfolio theory5Factor models8Asset allocation11Rebalancing3Portfolio construction7Benchmarks and tracking error5Performance measurement8Risk management6Fixed income and LDI5Currency and global3Implementation and costs5Active versus passive6India markets7Brainteasers5Career and fit16

Level

AnyCoreIntermediateHard

Type

AnyTechnicalCaseMarket viewFitBrainteaser

1. 029Why is mean-variance optimisation so unstable in practice, and what do you do about it?Portfolio constructionHardsuperdayMulti-assetQuantitative asset management

   Say this

   Because it is an error maximiser. The optimiser takes your most uncertain input, expected returns, and deliberately loads up on whichever asset has the highest estimate, so estimation error gets amplified rather than diversified. Small input changes produce enormous weight changes.

   Then walk it

   1. The mechanics of the failure: the solution involves inverting the covariance matrix, and with correlated assets that matrix is near-singular, so tiny differences in expected returns produce huge long-short positions in similar assets.
   2. Scale of the problem: Michaud's error maximisation and Chopra and Ziemba's work put the damage from expected return error at roughly ten times that of variance error and a hundred times that of covariance error. So the input you know least is the one that matters most.
   3. Fix one, better inputs. Shrink expected returns toward the mean or toward equilibrium, shrink the covariance matrix, impose a factor structure, and use longer histories for the covariance and shorter for nothing.
   4. Fix two, change the objective. Minimum variance, maximum diversification and risk parity avoid expected returns entirely, which removes the worst input at the cost of implicitly assuming returns are proportional to risk.
   5. Fix three, constrain and resample. Position bounds and asset class ranges cap the damage. Resampled efficiency, running the optimisation on many bootstrapped input draws and averaging the weights, produces far more stable portfolios than the point solution.
   6. Fix four, start from the market. Black-Litterman reverse-engineers the returns implied by market weights and then only tilts where you have a view with a confidence attached. That is the cleanest answer because it makes the default a sensible portfolio rather than a corner solution.
   7. The practical endpoint: most serious multi-asset houses run an optimiser as a diagnostic and a sense check, and set the actual policy mix with judgement, constraints and scenario testing. Saying that is more credible than claiming you trust the solver.

   Where candidates lose it

   Saying 'garbage in, garbage out' and leaving it there. The interviewer wants the specific mechanism, error maximisation via matrix inversion, the relative importance of the inputs, and at least two named remedies. Naming Black-Litterman or resampling turns a textbook answer into a practitioner's one.

   Expect next

   • Explain Black-Litterman then.
   • Why does minimum variance behave better out of sample?
   • What does resampling actually do to the weights?
2. 030Explain Black-Litterman and what problem it solves.Portfolio constructionHardsuperdayMulti-assetQuantitative asset management

   Say this

   It fixes the expected return input. Instead of forecasting returns from scratch, you reverse-engineer the returns implied by market capitalisation weights, treat those as the neutral prior, then blend in your own views with an explicit confidence. The output is a portfolio that only deviates where you actually have a view.

   Then walk it

   1. Step one, reverse optimisation. Take market weights, the covariance matrix and a risk aversion parameter, and solve backwards for the expected returns that would make market weights optimal. Those are the equilibrium returns.
   2. Step two, state views as portfolios, not as point forecasts of everything. A view can be absolute, 'EM equity returns 7 percent', or relative, 'European equities beat Japanese by 2 percent', and you attach a confidence, effectively a variance, to each.
   3. Step three, Bayesian blend. The posterior expected returns are a precision-weighted average of the equilibrium prior and your views. High confidence pulls the posterior toward your view; low confidence leaves it near equilibrium.
   4. Step four, optimise on the posterior. Because you started from market weights, no view means you end up at market weights, and one view produces a tilt concentrated in that view rather than a corner solution across twenty assets.
   5. Why that matters practically: it converts 'I am mildly bullish Japan' into a defined, sized, explainable deviation. Every position in the output can be traced to a view or to equilibrium, which is enormously easier to govern than an unconstrained optimiser's output.
   6. The limitations, which I would name: the tau and confidence parameters are subjective and the answer is sensitive to them, market cap weights as equilibrium are questionable for asset classes like bonds and private markets, and it still needs a covariance matrix. It makes the fragile input manageable rather than removing it.

   Where candidates lose it

   Describing it vaguely as 'combining views with the market'. The two mechanical steps that must be there are reverse optimisation from market weights to get the prior, and views expressed with a confidence that determines how far the posterior moves. Also concede that tau is a fudge factor, because interviewers who have implemented it know it is.

   Expect next

   • How do you set the confidence on a view?
   • What is the equilibrium portfolio for a bond allocator?
   • How is this different from just constraining the optimiser?
3. 033Your analysts give you return forecasts with wildly different levels of confidence. How do you build the portfolio?Portfolio constructionHardcase studyMulti-assetFundamental asset management

   Say this

   Make the confidence an input rather than a footnote. Convert each forecast into a signal with a dispersion attached, scale positions by the ratio of the expected return to its uncertainty, and let low-confidence views sit near benchmark weight instead of arguing them down qualitatively.

   Then walk it

   1. First, standardise the forecasts so they are comparable. Analysts express things differently, so I would convert everything to expected excess return over the same horizon, then to a z-score within the coverage universe.
   2. Then attach dispersion. Ask each analyst for a bear and bull case, not just a target, and use the spread as the uncertainty estimate. Position size then scales with expected return divided by variance, which is the Black-Litterman intuition applied at the stock level.
   3. Then adjust for track record rather than confidence expressed. Confidence and accuracy are barely correlated, and the analyst who sounds most certain is often the most overfitted. If I have hit rate data by analyst and by sector, I would shrink each forecast toward zero in proportion to their historical noise.
   4. Then handle correlation between views. Five high-conviction calls that all require the same rate path are one position. I would run the proposed portfolio through a factor model before trading, and cut the aggregate exposure rather than any single name.
   5. Then cap the damage. A hard maximum active weight regardless of stated conviction, because the largest single loss in a fundamental book usually comes from the position everyone agreed about.
   6. The organisational part matters too: if analysts learn that stated confidence drives sizing, confidence inflates. So the sizing rule should use the bear case and the historical accuracy, which are harder to game than a stated conviction score.

   Where candidates lose it

   Answering 'size by conviction' without saying how conviction becomes a number, or ignoring that stated confidence is gameable and uncorrelated with accuracy. The strong answer uses the bear case as the uncertainty measure, shrinks by track record, and aggregates the views through a factor model before trading.

   Expect next

   • How would you measure an analyst's hit rate?
   • What if the highest conviction ideas are all correlated?
   • Would you ever override the sizing rule?

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.