Multinomial Distribution in Finance

Table of Contents

    The multinomial distribution answers a question that a default/no-default model leaves open: how many borrowers improve, stay in the same rating, deteriorate or default? It assigns probabilities to complete combinations of category counts. That makes it useful for understanding categorical scenarios, provided the assumptions behind those probabilities are explicit.

    This guide develops the count model, works through a hypothetical rating-outcome example, and explains its mean, variance and negative count covariance. It also separates the distribution from a migration matrix, a trinomial pricing tree and a regression model. For the broader distinction between discrete counts and continuous financial variables, start with probability distributions in finance.

    What Is the Multinomial Distribution?

    One Trial Produces a Category; Many Trials Produce Counts

    A categorical trial has one outcome from a specified set of categories. The multinomial distribution describes the vector of category counts across a fixed number of independent trials with the same category probabilities.

    Write the count vector as X = (X1, …, Xk), where k is the number of categories. If there are n trials, the counts must satisfy X1 + … + Xk = n. Each count is a nonnegative integer. The probability vector p = (p1, …, pk) contains one probability per category, with each probability between zero and one and their sum equal to one. [1]

    The distinction matters. “Borrower 4 defaults” identifies one trial’s category. “One borrower defaults, one improves, seven remain unchanged and one deteriorates” identifies an entire count vector. It does not specify which borrowers occupy those categories.

    When Does the Count Model Apply?

    The elementary multinomial model requires the following conditions. MathWorks’ distribution documentation emphasizes independent trials, mutually exclusive outcomes and fixed category probabilities. [2]

    • A fixed number of trials. Specify n before counting the outcomes.
    • One exhaustive, mutually exclusive classification. Every trial belongs to exactly one category; no observation is omitted or counted twice.
    • The same probability vector for every trial. The model uses one common p, rather than different probabilities for different borrowers or dates.
    • Independence across trials. Knowing one trial’s outcome does not change the probabilities for the others.

    “Independent and identically distributed,” often shortened to iid, combines the last two conditions. It does not say that the resulting category counts are independent. Their fixed total creates a dependence that we will calculate below.

    The Multinomial Probability Formula

    The probability mass function (PMF) gives the probability of one particular count vector x = (x1, …, xk): [1]

    Multinomial PMF
    P(X1 = x1, …, Xk = xk) = [n! / (x1! × … × xk!)] × p1x1 × … × pkxk
    Valid counts are nonnegative integers totaling n. The factorial coefficient counts the distinct assignments that produce the same category totals.

    The product of probability powers describes one particular assignment of outcomes to the labelled trials. The multinomial coefficient, n! divided by the product of the count factorials, counts how many assignments have those totals. Adding their equal probabilities produces the count-vector probability.

    Here, “!” means factorial, with 0! = 1. A vector whose counts do not total n is impossible. A category with probability zero cannot have a positive count; when its count is also zero, its contribution to the product is treated as one.

    Example: Four Rating Outcomes

    Suppose an analyst studies ten different obligors, all initially rated BBB, over one year. For this teaching example only, assume their outcomes are independent and have the same probabilities below. These are invented inputs, not historical migration rates, current forecasts or a claim that real BBB obligors share identical risk.

    Classify each obligor once: a default during the year takes priority and remains in the default category for this exercise. For an obligor that does not default, compare its year-end rating with its initial rating. Assume every outcome is observed, with no rating withdrawals or missing cases. “Downgrade” below therefore excludes default.

    Hypothetical one-year outcomes for each initially BBB obligor
    Category Assumed probability Count being evaluated
    Upgrade 10% 1
    Unchanged 70% 7
    Nondefault downgrade 15% 1
    Default 5% 1
    Total 100% 10

    The question is the probability of exactly this full split, in any assignment across the ten obligors. First count the assignments; then weight them by their probability.

    Calculate the Probability of (1, 7, 1, 1)

    Assignments = 10! / (1! × 7! × 1! × 1!) = 720.

    Probability per assignment = 0.10 × 0.707 × 0.15 × 0.05 = 0.000061765725.

    Joint count probability = 720 × 0.000061765725 = 0.044471322, or approximately 4.45%.

    This 4.45% concerns all four category totals simultaneously. It is not the probability of one default regardless of what happens to the other obligors.

    What If We Only Care About Defaults?

    Combine upgrade, unchanged and nondefault downgrade into a single “no default” category. Its probability is 0.10 + 0.70 + 0.15 = 0.95. The default count D is then binomial under the same iid assumptions:

    P(D = 1) = 10 × 0.05 × 0.959 = 0.31512470, or approximately 31.51%. This larger event includes many different splits among the nine nondefault obligors. The earlier 4.45% event is just one of them.

    Every individual category count has a binomial marginal distribution, even though the counts jointly follow a multinomial distribution. A binary probability of default calculation therefore discards information about the other rating outcomes.

    Expected Counts, Variance and Covariance

    For a category with probability pi, the expected count is n times that probability. Its variance measures how the count varies across repetitions of the assumed experiment. For two different categories, the covariance is negative when both probabilities are positive. [2]

    Multinomial Count Moments
    E[Xi] = npi
    Var(Xi) = npi(1 − pi)
    Cov(Xi, Xj) = −npipj, for i ≠ j
    Expectations are counts; variances and covariances have units of count squared.

    Keep n = 10 and the same four hypothetical probabilities. The expected counts and their variances are:

    Moments of the hypothetical ten-obligor count vector
    Category Expected count: 10p Variance: 10p(1 − p)
    Upgrade 10 × 0.10 = 1 10 × 0.10 × 0.90 = 0.900
    Unchanged 10 × 0.70 = 7 10 × 0.70 × 0.30 = 2.100
    Nondefault downgrade 10 × 0.15 = 1.5 10 × 0.15 × 0.85 = 1.275
    Default 10 × 0.05 = 0.5 10 × 0.05 × 0.95 = 0.475

    An expectation of 0.5 defaults is a long-run model average across repetitions. Every realized default count remains an integer. Likewise, an expected count vector is not a guaranteed outcome, and it need not be a possible integer vector.

    Why Are Different Category Counts Negatively Related?

    Each obligor can occupy only one category. Within a single trial, the indicators for “upgrade” and “unchanged” cannot both equal one. Their product is zero, making their covariance 0 − (0.10 × 0.70) = −0.07. Independence across the ten obligors lets us add those ten contributions:

    Cov(Xupgrade, Xunchanged) = −10 × 0.10 × 0.70 = −0.70. The category counts compete for the same fixed total. For a broader explanation of covariance’s units and interpretation, see correlation and covariance.

    This does not imply that one obligor’s default makes another obligor safer. The negative covariance relates different category totals in the iid experiment. Dependence between defaults of different obligors is a separate modeling question.

    Why Shared Credit Factors Break the Iid Example

    Two obligors may respond to the same recession, commodity price or regional shock. Basel’s credit-concentration discussion explicitly identifies common or correlated risk factors, including sector and geographic exposures. [5] A common initial rating does not eliminate those relationships.

    For a default count, dependence adds cross-obligor covariance terms to the variance. Positive covariance between different obligors’ default indicators raises count variance relative to the iid calculation with the same marginal default probabilities. The elementary value of 0.475 in our example excludes those terms; it is not a validated portfolio risk estimate.

    Conditional Independence Is a Different Assumption

    A factor model might assume outcomes are independent after conditioning on a common economic state. If that state changes the category probabilities, averaging across states generally produces a mixture rather than the fixed-probability multinomial model. Substituting average probabilities into the simple PMF does not generally recover the joint distribution.

    Heterogeneity is another problem: independent obligors with different probability vectors do not generally produce the elementary multinomial count law either. Independence and identical probabilities must each be justified. Monte Carlo simulation can implement richer assumptions, but simulation itself does not establish that those assumptions are right.

    Multinomial Counts vs a Rating Migration Matrix

    A rating migration matrix organizes destination probabilities by initial rating. A row for BBB describes the possible destinations of one obligor starting at BBB over the stated horizon. MathWorks explicitly distinguishes standard rating rows from issuer-specific probability rows. [3]

    Grouping that BBB row’s destinations into upgrade, unchanged, nondefault downgrade and default can supply a probability vector like the one in our example. Moving from a one-obligor row to a distribution of portfolio counts requires the additional common-probability and independence assumptions. The matrix alone does not specify the joint behavior of different obligors.

    The observation convention also matters. A cohort transition estimate compares ratings at specified endpoints, so intermediate rating changes can be absent from its counts. MathWorks contrasts this with duration estimation, which uses the dated rating history. [4] An obligor that is downgraded and then returns to BBB can be “unchanged” in an endpoint classification.

    Choose the object that answers the question
    Object Question answered Key distinction
    Categorical distribution Which category occurs on one trial? One outcome; no aggregate count calculation yet.
    Binomial distribution How many trials fall into one category versus all others? One count; the complementary count is n minus it.
    Multinomial distribution What is the probability of a full vector of counts? Several category counts sharing the same fixed total.
    Migration matrix Where can an obligor move from each initial rating? Conditional destination probabilities, not a cross-obligor dependence model.
    Multinomial logit How do explanatory variables predict category probabilities? A regression specification, rather than just a supplied fixed probability vector.
    Ordered logit How do explanatory variables relate to ordered outcome categories? Uses ordering and cutpoints, with additional structural assumptions.

    Multinomial logit can estimate probabilities from borrower characteristics instead of treating p as given. Ordered logit is a candidate when the categories have a meaningful ranking, as rating grades do. The standard ordered-logit specification uses a proportional-odds structure; ordered labels do not imply equally spaced numerical ratings. [6] [7]

    These regression models address how probabilities vary with predictors. They do not, by themselves, establish independence across obligors. For the two-outcome starting point, see binary logit and probit models.

    How Does This Relate to a Trinomial Tree?

    Trinomial means three categories in the count setting. A trinomial tree likewise has top, middle and bottom branches, but it retains the sequence of branch choices through time. MathWorks’ tree documentation explicitly describes paths by their ordered branch sequence. [8]

    For a separate hypothetical example, assume three independent steps with fixed probabilities of 25% up, 50% flat and 25% down at every step. One particular order, up-flat-down, has probability:

    0.25 × 0.50 × 0.25 = 0.03125 = 3.125%. Exactly one outcome of each kind can occur in 3! / (1! × 1! × 1!) = 6 orders, so the count-vector probability is 6 × 0.03125 = 0.1875 = 18.75%.

    The count vector has discarded the order. A tree can preserve it, which matters when a payoff depends on the path. Moreover, actual lattice probabilities may depend on time or state, so three branches alone do not establish an iid multinomial model. The numbers here are arbitrary category probabilities, not calibrated risk-neutral pricing inputs. For replication, discounting and option valuation, see the separate binomial option pricing model guide.

    How to Evaluate a Multinomial Model

    1. Define one trial and one horizon. Distinguish an obligor from a bond position; several bonds issued by the same company are not several independent company outcomes.
    2. Write the classification rules. Resolve default, withdrawal, missing observations and intermediate rating changes before counting.
    3. Identify where probabilities came from. Separate assumed teaching values, historical frequency estimates and conditional model forecasts. Check every probability and their total.
    4. Challenge both iid conditions. Ask whether borrowers have comparable probabilities and whether shared factors create dependence.
    5. Specify the event precisely. A complete count split, exactly one default and at least one default are different events.
    6. Translate counts into the actual risk question. Dollar losses also require exposures, recoveries and, for nondefault migrations, valuation assumptions. Counts alone do not determine portfolio loss.

    Common Mistakes

    • Omitting the factorial coefficient. The probability product describes a particular assignment; a count event generally includes many assignments.
    • Multiplying marginal binomial probabilities. The category counts are dependent. Use the joint PMF for the full vector.
    • Overlapping downgrade and default. An obligor cannot be counted in both categories in this multinomial classification.
    • Treating expected counts as realized counts. A fractional expected default count is an average, not a fractional observed default.
    • Confusing two kinds of covariance. Negative covariance across category totals does not justify assuming independent or negatively correlated defaults across obligors.
    • Using a transition row as a complete portfolio model. Marginal destination probabilities leave cross-obligor dependence unspecified.

    Limitations of the Multinomial Distribution

    The distribution is exact for its stated experiment, but fitting a financial problem into that experiment can be difficult. Probabilities may differ across borrowers, change over time or respond to shared shocks. An estimated probability vector also carries estimation uncertainty, which the elementary known-p calculation does not include.

    Rare outcomes create a further practical limit: an observed zero count does not establish that the true probability is zero. More categories can preserve useful detail while leaving fewer observations in each category. Combining them simplifies estimation but can hide distinctions that matter for losses or decisions.

    Use the multinomial distribution as a transparent baseline for categorical counts. Whether it is adequate for a particular portfolio depends on evidence about the probabilities, classification and dependence structure.

    Frequently Asked Questions

    Binomial models one success count across fixed independent trials with a common success probability. Multinomial models the full vector of category counts under the corresponding assumptions. With two categories, the second count is n minus the first, so the multinomial reduces to the binomial case.

    Generally no. Trials are assumed independent, but category counts share a fixed total. For different categories, their covariance is minus n times the product of their probabilities, which is negative when both probabilities are positive. A zero-probability category has a constant zero count.

    No. Its rows describe destination probabilities conditional on an initial rating. A multinomial count model for several obligors additionally needs common category probabilities and independence across those obligors. Shared economic exposures and issuer differences can violate those assumptions.

    No. The distribution assigns probabilities to counts for given inputs n and p. Multinomial regression models how category probabilities depend on explanatory variables. Ordered regression can use a meaningful ordering of categories, subject to its additional assumptions. Neither label alone specifies cross-observation dependence.

    Yes. Different categories can have different probabilities, as the rating example shows. What the elementary model requires is that the entire probability vector stays the same across trials. If one obligor has a different vector from another, their aggregate counts do not generally follow that elementary multinomial law.

    Sources

    1. SciPy: scipy.stats.multinomial, v1.18.0 documentation. Probability mass function, support and the two-category relationship.
    2. MathWorks: Multinomial Distribution, R2026b documentation. Trial assumptions, expected counts, variance and covariance.
    3. MathWorks: Convert from Credit Quality Thresholds to Transition Probabilities. Rating-row and issuer-specific destination probabilities.
    4. MathWorks: Estimate Transition Probabilities from Credit Ratings Data. Cohort endpoint observations versus duration estimation.
    5. Basel Committee on Banking Supervision: SRP32, Credit Risk, paragraphs 32.8–32.9. Common and correlated factors in credit concentrations; version effective January 1, 2023.
    6. statsmodels: Multinomial Logit Model documentation. Categorical responses and explanatory variables.
    7. statsmodels: OrderedModel documentation. Ordered labels, latent cutpoints and the proportional-odds logistic specification.
    8. MathWorks: Entries from a Node of a Recombining Trinomial Tree. Ordered top, middle and bottom branch sequences.

    Sources accessed November 14, 2026. All numerical scenarios are independently calculated hypothetical examples, not observed issuer statistics or investment results.

    Disclaimer

    This article is for educational and informational purposes only and does not constitute investment advice. Its assumed rating probabilities and categorical scenarios are not forecasts or a validated credit-risk model. Actual credit outcomes, losses and valuations depend on issuer characteristics, shared risk factors, exposures, recoveries and model assumptions. Consult a qualified financial professional about your circumstances.