Green Book Quant Guide: Chapters, Study Plan and Solved Practice

The quant Green Book is usually A Practical Guide to Quantitative Finance Interviews, written by Xinfeng Zhou and edited by Brett Jiu. Its green cover supplies the informal name. The 2008 Google Books record identifies the title, author, publisher and subject coverage. Google Books bibliographic record.

The useful purchase question is whether you need a compact collection of worked interview problems across several mathematical subjects. If you already know the underlying mathematics but struggle to choose a method, explain a derivation or solve under observation, this format can help. If conditional probability, differentiation and basic matrices are new to you, a foundations course should come first.

Use the book as a workbook. Cover an answer, write the assumptions and attempt a solution. Then check the argument, including intermediate steps the printed answer may compress. The benefit comes from that effort; owning the book or reaching the last page does not establish interview readiness.

For a quick next step, try the prerequisite diagnostic, choose the chapters for your role, or start with the original solved practice. The eight-week plan turns those choices into a schedule.

There is no meaningful universal percentage of an interview that this book covers. A trading problem, a statistical research discussion and a C++ systems interview require different preparation. Start with the employer’s current candidate information and use this guide to connect those requirements to the relevant material.

Book listings disagree about pagination and sometimes attach misleading edition labels. Identify the copy through several fields together.

Scroll horizontally to read every column.

Sources: 2008 Lulu record, 2008 CreateSpace catalog, 2020 paperback catalog, and KAIST holding, checked October 3, 2026.

For a new purchase, compare the seller’s ISBN and a photograph of the copyright page. For a used copy, confirm that the contents, answer pages and index are present. A difference between numbered text pages and total printed pages can explain some discrepancies, but these records alone do not prove the reason for every mismatch.

Our chapter map follows the catalogued first-edition structure. We have not verified a complete line-by-line list of changes between editions, so we do not promise that the 2020 printing fixes every error or adds a particular number of questions. This plan uses topics rather than page numbers so it remains useful across copies.

Some reviews say “six chapters” when they mean six technical areas. The catalog includes an introductory chapter before those areas. The following is a paraphrased coverage map, not a reproduction of the book’s contents. Catalogued table of contents.

Scroll horizontally to read every column.

The last column is our proposed evidence of learning. It is more informative than a tick against a chapter. For example, an optimisation answer is incomplete if it finds a stationary point without checking that it is a minimum. A pricing answer is incomplete if it produces a number that violates a payoff bound.

Read selectively within a chapter as well as between chapters. A probability mistake caused by confusing independent trials with unconditional independence needs different repair from a counting mistake. Likewise, matrix fluency and Brownian-motion fluency are separate skills. Record the method that failed so your next session addresses a specific gap.

You do not need a particular degree title to use the book. You do need enough mathematical fluency to reconstruct an argument. Try these checks without searching for a formula, then explain the answer aloud.

Scroll horizontally to read every column.

Treat these as placement checks, not a validated assessment or an interview pass score. If the first two are comfortable, begin probability problems while repairing other gaps. If you need the answers to understand what the questions mean, spend one or two preliminary weeks learning the definitions and solving simpler exercises. Continue longer if needed.

For probability, build from events to conditioning, random variables, expectation and variance. For calculus, practise derivatives, integrals and boundary checks. For linear algebra, start with systems, vectors and covariance matrices. For chapter 5, a discrete-state recurrence is a more accessible first step than memorising Itô’s lemma.

The free conditional-probability lesson and free Python starter course provide useful entry points. They complement the book’s topics; they do not reproduce its problem set.

Role labels are broad. A researcher developing equity signals can need different mathematics from a researcher pricing exotic derivatives. Use the job description, recruiter guidance and official candidate material to decide the emphasis. If those distinctions are unfamiliar, our guide to what quants do explains the work and deliverables before you choose a reading priority.

Scroll horizontally to read every column.

These are study recommendations, not measured interview-frequency statistics. Avoid the rule that every researcher must prioritise stochastic calculus over statistics. The right ordering follows the work.

One concrete example is Jane Street’s official candidate page. It describes trading interviews as collaborative problem solving, says finance knowledge is not tested for that discipline, and describes research interviews as combining trading and software-engineering elements. It also links its own probability and markets preparation guide. That supports choosing probability and coding preparation for those processes; it does not establish that Jane Street uses Zhou’s questions. Jane Street interview guidance.

For a bank pricing role, finance chapters may be central. For a developer role focused on infrastructure, studying option Greeks for weeks while avoiding coding can be a poor allocation. Ask what you will need to demonstrate, then reserve time for that demonstration.

Its main advantage is breadth in a compact problem format. You can expose a weakness in conditional probability, matrix algebra or recursive reasoning without committing to a separate large textbook for each topic. Short worked solutions can also prompt useful reconstruction: fill every omitted algebra step and justify each assumption yourself.

That compression creates limits. A solution can look obvious after you read it while remaining impossible to produce from a blank page. If you cannot define the state, choose the conditioning event or explain the payoff identity, more reading of the same paragraph may not resolve the underlying gap.

Chapter 7 supplies algorithmic and numerical material, so calling the book “a book with no coding” would be inaccurate. It still cannot provide the volume of implementation, debugging and live communication needed for a coding interview. Run your solutions and test inputs that could break them.

For statistical research, add a task where the answer is not a single exact number. Define a hypothesis, choose a baseline, prevent future information entering training, evaluate uncertainty and explain a failed result. For trading, add decisions made with incomplete information when the target process asks for them. For pricing, add numerical error, model assumptions and payoff checks.

An older mathematical workbook can remain useful because Bayes’ rule and no-arbitrage identities still work. Whether its problem mix matches your next interview is a separate question. Treat “essential for everyone” and “enough to pass” as claims that require evidence, rather than buying rules.

Choose a supplement for the gap you actually found. The comparison below uses catalog and publisher descriptions for scope; the recommendations are our editorial judgment. We do not score books from unverified claims about offer rates or question overlap.

Scroll horizontally to read every column.

Sources: Crack’s author-published catalog description, Joshi and co-authors’ publisher description, and Dover’s book listing.

If money or time is limited, avoid buying all four immediately. Solve a varied set in your first book, identify the missing skill and inspect a sample of the proposed supplement. Someone weak in foundational probability may benefit more from a structured introductory text than from a harder puzzle collection. Someone weak in Python needs implementation practice, even if another finance book has excellent solutions.

Prices, availability and shipping depend on edition and location. Compare the total purchase cost at checkout. We have omitted approximate price tables that could confuse an old used printing with a new edition.

This is a planning template for someone who already knows the prerequisites, with roughly five to seven hours available per week. It is not an observed average completion time or a guarantee that you will finish every problem. Add weeks for foundation repair and compress areas you can already demonstrate.

Scroll horizontally to read every column.

For each ordinary session, allocate about 10 minutes to recalling an earlier method, 25 minutes to a fresh attempt, 15 minutes to reconstructing and checking the solution, and 10 minutes to an altered question. Those are adjustable time boxes. If a concept is unfamiliar, spend the attempt defining the model and identifying what you need to learn.

At three hours per week, spread this sequence across more weeks rather than deleting independent attempts. With two weeks until an interview, use the diagnostic to select relevant problems and practise the actual format; do not attempt to learn every chapter from scratch. With a strong background, move more time from familiar derivations to mixed questions and communication.

Progress means you can recognise the mathematical structure of a fresh problem, explain your plan and recover when an assumption changes. Counting completed pages measures a different thing.

Use four passes: attempt, reconstruct, vary and retest. On the first pass, write what is known and what is random. On the second, compare your argument with the answer and fill missing steps. On the third, change a parameter or assumption. On the fourth, return from a blank page after a delay.

Keep an error log that names the mathematical failure.

Scroll horizontally to read every column.

Record whether you solved independently, used a hint or reconstructed a shown answer. All three can contribute to learning, but they represent different evidence. “Careless” is too vague: write “omitted the covariance term” or “counted unordered pairs as equally likely”.

When speaking, explain the route before the arithmetic. A useful order is assumptions, model, calculation, check and limitation. If stuck, discuss a small case or a bound. These steps reveal useful reasoning even before you have a final answer.

The next exercises were written for this guide. They are not copied Green Book questions, an official solutions manual or verified questions from any employer. They illustrate the mathematical families the book covers.

Try each before reading its solution. A correct final number is one part of the task. Also state the assumptions, justify the method and perform the independent check. The probabilities, optimisation values, recurrence and payoff identities below were recalculated with exact arithmetic or explicit checks on October 3, 2026.

Use the first three to assess reasoning and probability, the next three for optimisation and random processes, and the last two for finance and implementation. You can choose an order by role. If a question reveals a prerequisite gap, repair that gap before adding harder variants.

Problem. Exactly one of ten sensors has failed. A test can inspect any chosen group and returns positive if that group contains the failed sensor. Tests have no errors and can run in parallel. What is the minimum number of tests that identifies the sensor in the worst case? Give a construction.

Solution. Each test returns one of two outcomes. Three tests give at most 2³ = 8 result patterns, which cannot distinguish ten possible failed sensors. Therefore at least four tests are necessary.

Four suffice. Label sensors 1 through 10 with four-bit binary numbers: 0001, 0010, and so on through 1010. Test 1 includes sensors whose first bit is 1, test 2 those whose second bit is 1, and likewise for the remaining bits. Because exactly one sensor failed, the four results reproduce that sensor’s label.

Check. Every sensor has a unique result pattern. Sensor 10, labelled 1010, produces positive, negative, positive, negative in the same bit order. The lower bound and construction agree, so four is optimal.

What changes? With two failed sensors, pooled positives combine their bits and may collide with another pair’s pattern. With false negatives, unique noiseless labels are insufficient. The argument relies on the stated model, not merely on knowing the binary formula.

Problem. A coin has heads probability 3/4 with prior probability 1/3, or heads probability 1/4 with prior probability 2/3. Conditional on the selected coin, flips are independent. You observe two heads. What is the probability that the next flip is heads?

Solution. First update the probability of the coin type. The unnormalised weight for the high-heads coin is (1/3)(3/4)² = 3/16. The low-heads weight is (2/3)(1/4)² = 1/24. Their sum is 11/48. The posterior high-heads probability is therefore (3/16)/(11/48) = 9/11.

The next-flip probability averages the two heads rates using the posterior weights:

P(next H | HH) = (9/11)(3/4) + (2/11)(1/4) = 29/44.

Check. Before observing flips, the heads probability was 5/12. Two heads should increase belief in the high-heads coin, and 29/44 is greater than 5/12 while remaining below 3/4. The observation is evidence, not certainty.

Common error. The flips are independent given the coin type. They are not independent when the hidden type is unknown. The first two flips inform that type, which changes the prediction. Change the prior to 1/2 and rebuild the calculation instead of memorising 29/44.

Problem. A first draw is uniform on 1, 2, 3 and 4. You can accept that amount or pay 0.25 to replace it with an independent second draw from the same distribution. You must accept the second draw. You maximise expected net payoff. What is the optimal rule and its expected value?

Solution. Work backward. The second draw has mean (1 + 2 + 3 + 4)/4 = 2.5. Replacing therefore has expected net value 2.5 − 0.25 = 2.25. After the first draw x, compare x with 2.25: replace 1 or 2, and accept 3 or 4.

The overall expected net payoff is:

(2.25 + 2.25 + 3 + 4) / 4 = 2.875 = 23/8.

Check. Always accepting gives 2.5. Always replacing gives 2.25. The conditional policy gives more than either, which is possible because it uses information from the first draw.

What changes? The fee is incurred only when replacing. Charging it on every path changes the problem. With fee c, compare x against 2.5 − c. A fee above 1.5 makes replacement unattractive for every possible first draw. At exactly 1.5, replacing a 1 and accepting it have equal expected value. This rule assumes risk neutrality; a preference for certainty can change the decision.

Problem. Allocate positive continuous amounts x and y of measurement effort with x + y = 12. A simplified uncertainty cost is 2/x + 8/y. Find the allocation that minimises cost and prove it is a minimum.

Solution. Substitute y = 12 − x, so 0 < x < 12. Let f(x) = 2/x + 8/(12 − x). Then:

f′(x) = −2/x² + 8/(12 − x)².

Setting the derivative to zero gives (12 − x)² = 4x². Both quantities are positive in the domain, so 12 − x = 2x. Hence x = 4 and y = 8. The cost is 2/4 + 8/8 = 1.5.

To establish a minimum, calculate f″(x) = 4/x³ + 16/(12 − x)³, which is positive throughout the domain. The cost also diverges near either boundary. The unique stationary point is the global minimum.

Check. Equal effort gives 2/6 + 8/6 = 5/3, which exceeds 1.5. The larger coefficient receives more effort, but its optimal effort is twice as large, not four times as large. This follows from the squared terms in the derivative.

What changes? If effort must be integer-valued, inspect nearby feasible allocations. If the cost model or minimum-effort constraint changes, derive the optimisation again. Symmetry is not available when the coefficients differ.

Problem. Two abstract return variables have variances 4 and 9 and covariance 2, all in the same squared units. A combination assigns weights w and 1 − w. Find the weight with minimum variance and check the covariance matrix is valid. This is a mathematical exercise, not a portfolio recommendation.

Solution. The covariance matrix is [[4, 2], [2, 9]]. Its first leading principal minor is 4 and its determinant is 36 − 4 = 32. Both are positive, so the matrix is positive definite.

The combined variance is:

V(w) = 4w² + 9(1−w)² + 2·2w(1−w) = 9w² − 14w + 9.

Differentiate: V′(w) = 18w − 14. The minimum occurs at w = 7/9, with the second weight 2/9. Since V″(w) = 18 > 0, this is a unique global minimum. It also lies in the long-only interval 0 ≤ w ≤ 1.

Substitution gives V(7/9) = 32/9, approximately 3.556.

Check. The result is below either individual variance, 4 and 9. A direct weighted-matrix calculation should give the same value. Omitting the covariance term would solve a different problem and produce a different allocation.

What changes? Unknown or estimated covariance introduces estimation uncertainty. A clean algebraic minimum does not establish future investment performance. Additional constraints can put the optimum on a boundary rather than at the unconstrained stationary point.

Problem. A token starts at state 1. At every step it moves up one with probability 2/3 and down one with probability 1/3, independently of earlier moves. It stops on reaching 0 or 4. What is the probability it reaches 4 first?

Solution. Let f(k) be the success probability starting at state k. The boundaries are f(0) = 0 and f(4) = 1. For k = 1, 2, 3, condition on the first move:

f(k) = (2/3)f(k+1) + (1/3)f(k−1).

The general solution of this difference equation is A + B(1/2)^k. Applying the two boundaries gives:

f(k) = [1 − (1/2)^k] / [1 − (1/2)^4].

Thus f(1) = (1/2)/(15/16) = 8/15, approximately 0.5333.

Check. Substitute the formula into each of the three interior recurrence equations and both boundaries. For a fair walk, the success probability from 1 would be 1/4. Positive upward drift should increase it; 8/15 does. The result remains between zero and one.

What changes? The fair formula k/4 cannot be used unchanged for a biased walk. If transition probabilities depend on the state, write separate equations using those probabilities and solve the resulting system. Defining the state and boundary conditions is the transferable skill.

Problem. A non-dividend-paying stock trades at 80. European call and put options have the same strike 82 and expiry. The discount factor to expiry is 0.96. The call costs 6. What put price is consistent with parity? If a put costs 6, describe the discrepancy using cash flows.

Solution. Under the frictionless assumptions of put-call parity:

C + discounted strike = P + S.

The discounted strike is 82 × 0.96 = 78.72, so P = 6 + 78.72 − 80 = 4.72. If P is 6, the put-plus-stock portfolio costs 86, while the call-plus-bond portfolio costs 84.72.

The latter is cheaper by 1.28. Buying the call and a bond that pays 82, while shorting the put and stock, receives 1.28 initially. At expiry the net payoff is:

max(S_T−82, 0) + 82 − max(82−S_T, 0) − S_T = 0.

Check. Verify that payoff separately for S_T below and above 82. The two portfolios pay the same amount in every state. The argument assumes you can borrow, short and trade at the quoted prices without costs or restrictions.

Critical pitfall. If the call input were 1 instead, parity would return P = −0.28. Do not clip that number to zero and call it a valid price. The call itself violates the lower bound max(80 − 78.72, 0) = 1.28. Check the inputs and assumptions.

The free option-pricing tool can support sensitivity practice. A calculator output still needs payoff and assumption checks.

Problem. Given returns [0.02, −0.01, 0.03, 0.00], calculate every complete three-observation trailing mean. Write an O(n) algorithm. If a mean is a forecast input for the next period, specify when it is available.

Solution. The first complete mean is (0.02 − 0.01 + 0.03)/3 = 1/75, approximately 0.013333. The next is (−0.01 + 0.03 + 0.00)/3 = 1/150, approximately 0.006667. Update a running sum by adding the newest observation and removing the observation that leaves the window.

def trailing_means(values, window):
    if isinstance(window, bool) or not isinstance(window, int) or window < 1:
        raise ValueError("window must be a positive integer")
    result = []
    total = 0.0
    for i, value in enumerate(values):
        total += value
        if i >= window:
            total -= values[i - window]
        if i >= window - 1:
            result.append(total / window)
    return result

sample = [0.02, -0.01, 0.03, 0.00]
actual = trailing_means(sample, 3)
reference = [sum(sample[i:i+3]) / 3 for i in range(2)]
assert all(abs(a-b) < 1e-12 for a, b in zip(actual, reference))
assert len(actual) == 2
assert trailing_means([0.2], 3) == []
assert trailing_means([0.2], 1) == [0.2]

Check and complexity. The reference uses separate slices, so it provides a different calculation path. The running algorithm visits each input once: O(n) time and O(1) additional working space apart from the returned output. It assumes finite numeric input and uses floating-point arithmetic.

Timing. The mean ending at time t can inform a decision after that observation is available. It cannot earn the return at t using information first observed at t’s close. Keep feature timing, execution timing and realised return separate. Test empty inputs, short series and invalid windows before extending the algorithm.

Memorising a final number. A remembered answer can survive a change in names or units but fail when the conditioning event changes. Redraw the sample space after any material change.

Skipping checks because the solution looks authoritative. Check probability bounds, dimensions, endpoints, feasible constraints and payoff identities. A negative option price is a signal to investigate assumptions or inputs. A stationary point does not automatically minimise a cost.

Treating every research role as a pricing role. Brownian motion may matter greatly for one team, while data leakage, inference and experiment design matter more for another. Use the actual requirements.

Calling reconstruction an independent solution. If you used a hint, record it. Later, try a fresh problem from the same method family. This makes progress visible without pretending every attempt was unaided.

Reading programming answers without running them. Implement the algorithm, state its complexity and choose edge cases. For numerical methods, compare with a simpler case whose answer you can calculate by hand.

Using all study time on comfortable chapters. Put a short repair session for your weakest relevant topic on the calendar. An extra hour of familiar puzzles may feel productive while leaving the main gap untouched.

Assuming completion guarantees an offer. Interviews can include coding, research discussion, collaboration and material outside any single book. Use a mock session to assess how you communicate when you do not immediately know the answer.

Use a legitimate bookseller or a library holding and match the title, author and ISBN. The catalog’s library lookup is a starting point for finding a copy; availability depends on your library.

If you search for “Green Book quant PDF”, search results can include full scans on unrelated websites. Search visibility alone does not establish that a copy is authorised or accurate. We do not offer a full-book PDF or reproduce its questions. For digital access, check the seller’s format and the publisher or rights holder’s authorisation.

Online walkthroughs can be useful when a step is unclear, but check whether they use the same assumptions and edition. Compare the reasoning rather than only the answer. If two explanations disagree, test a small case, derive a boundary condition or enumerate outcomes; do not decide by the page’s ranking.

This guide is independent of Zhou, the publisher and the other authors mentioned. It is published by a quant learning product, and its practice links are to that product. The chapter recommendations are editorial guidance, not a claim of author endorsement. The eight exercises above teach related methods through original examples and are not a replacement for the book’s full contents.

Pick a next step that tests the method you just studied.

Scroll horizontally to read every column.

The article and its complete worked examples are public. You can do the paper exercises and run the short code locally without buying a plan. Pro extends the guided learning and research-project workflow; it does not unlock an official Green Book solution set. Consult the current plan details if that additional workflow addresses a gap you have identified.

For a research role, aim to finish with a small piece of work you can defend: inputs, assumptions, code, checks, limitations and a concise explanation. That gives the mathematics a use beyond recognising a printed solution.