How Markov-Functional Interest-Rate Algorithms Fit Smile Distributions with a Low-Dimensional State: Numeraire Mapping, Digital Swaptions, Backward Induction, Calibration and Dynamics Failure

Reader question: Can an interest-rate model remain low-dimensional enough for fast Bermudan-style pricing while still matching an entire market-implied caplet or swaption smile rather than only one volatility number?

Markov-functional interest-rate models solve this by separating the state process from the market-observable rate mapping. A low-dimensional Markov process drives the model, while discount bonds, forward rates or swap rates are defined as functions of that state. The functional mapping is then calibrated so that chosen market option prices—or equivalently an implied terminal rate distribution—are reproduced.

This article owns one precise page role: how a low-dimensional Markov state is mapped into market-implied interest-rate distributions for derivative pricing. It does not own HJM drift restrictions, Hull–White short-rate calibration, LIBOR Market Model forward-rate dynamics, Black/Bachelier smile quotation or Jamshidian decomposition. Those are neighbouring canonical roles already represented in the Bukit Timah Tutor finance-and-banking-algorithms estate.

This is public mathematical and computational education. It is not financial advice, a trading recommendation or a claim that exact calibration of vanilla options guarantees correct exotic-option dynamics.

1. Why one-factor short-rate models can be too rigid

A traditional one-factor short-rate model specifies a stochastic short rate such as:

dr_t = [θ(t) − ar_t]dt + σ dW_t.

Once the short-rate process is fixed, bond and swaption distributions follow from that process.

This is elegant, but one low-dimensional diffusion may have difficulty matching a full market smile or skew across strikes while also retaining tractable dynamics.

Markov-functional modelling reverses the emphasis. Instead of forcing the market rate distribution to emerge from a particular short-rate formula, it chooses a tractable Markov driver and then constructs a function that maps that driver into the market rate distribution.

2. The basic Markov-functional statement

Let X_t be a low-dimensional Markov process. In a one-factor version, X_t is scalar.

A discount bond can be written as:

P(t,T) = P(t,T; X_t).

Likewise, a swap rate or forward rate observed at time t becomes a deterministic function of X_t:

S(t) = S(t, X_t).

The process X contains the stochastic dimension. The function S(·) contains the calibrated market mapping.

3. Markov does not mean Gaussian rates

The driving state X can be chosen to have simple dynamics—often Gaussian or close to Gaussian—while the mapped rate S(X) can have a highly non-Gaussian distribution.

A nonlinear monotone mapping can transform a normal state into a skewed or heavy-tailed swap-rate distribution.

This is one of the model’s main attractions:

simple state dynamics + flexible rate distribution.

4. Numeraire and martingale ratios

Interest-rate pricing becomes simpler under a suitable numeraire N_t.

Under the measure associated with that numeraire, the ratio:

P(t,T)/N_t

is a martingale.

Therefore:

P(t,T)/N_t = E[P(u,T)/N_u | F_t]

for t ≤ u, under the chosen pricing measure.

Markov-functional algorithms exploit this relationship to propagate bond-value functions backward through exercise/calibration dates.

5. Why a swap rate can be the natural market object

For a fixed-start swap with payment dates T₁,…,T_m, the forward swap rate can be written:

S(t) = [P(t,T₀) − P(t,T_m)] / A(t),

where the swap annuity is:

A(t) = Σ δ_iP(t,T_i).

European swaptions are quoted directly on this rate. A Markov-functional swaption model can therefore calibrate the mapping X → S so that the market’s swaption smile is reproduced at the option expiry.

6. The monotone mapping idea

Suppose at option expiry T_n the state X_n has a known conditional distribution under a convenient measure.

Assume the model swap rate is monotone in the state:

S_n = f_n(X_n).

If f_n is increasing, then:

P(S_n ≤ s) = P(X_n ≤ f_n⁻¹(s)).

This gives a quantile-mapping interpretation:

market-implied swap-rate quantile ↔ state-process quantile.

The calibrated function f_n can therefore make the model swap-rate distribution match a distribution inferred from market option prices.

7. Digital options reveal the distribution

A digital swaption pays when the terminal swap rate crosses a strike. Under the appropriate annuity numeraire, its price is directly related to a tail probability of the swap rate.

If a smooth vanilla swaption smile is available, digital prices can be inferred from strike derivatives of vanilla option prices.

Schematically:

digital probability at strike K ≈ derivative of option price with respect to K.

This connects Markov-functional calibration with the same state-price logic that underlies Breeden–Litzenberger risk-neutral density extraction.

8. Smile input is usually modelled through Black, Bachelier or another quote convention

Market swaption quotes are commonly supplied as normal or lognormal implied volatilities, sometimes with shifts or smile parameterisations.

The Markov-functional engine first needs a smooth arbitrage-consistent option-price curve across strike.

See Black-76 and Bachelier swaption algorithms for the quotation-to-price layer.

9. Terminal calibration is only the first step

Matching the distribution of a terminal swap rate at one expiry is not enough to price a multi-exercise Bermudan product.

The model also needs consistent bond-price functions at earlier dates.

This is where backward induction and the Markov property become important.

10. Backward induction of bond functions

Suppose a discount bond P(T_n,T) has already been represented as a function of X_n.

At the previous date T_n−1:

P(T_n−1,T)/N(T_n−1) = E[P(T_n,T)/N(T_n) | X_n−1].

Because X is Markov, this conditional expectation depends only on the current state X_n−1, not the entire path history.

Numerical quadrature, lattice methods or conditional Gaussian integration can therefore convert the later bond function into an earlier bond function.

11. Calibrate, then step backward

A stylised one-factor swaption Markov-functional algorithm proceeds:

1. Choose a Markov state process X.
2. Build a time grid at relevant swaption/Bermudan dates.
3. Choose a numeraire/measure.
4. At the latest calibration date, infer the market swap-rate distribution from the smile.
5. Construct the monotone state-to-swap-rate mapping.
6. Recover discount-bond functions consistent with that swap rate.
7. Use conditional expectations to propagate bond functions one step backward.
8. At the next calibration date, impose the next market smile/distribution.
9. Continue backward until today.
10. Price exotics using the resulting low-dimensional dynamic state.

12. The model can fit vanilla distributions exactly—subject to the chosen interpolation

A central appeal of the Markov-functional framework is that, with appropriate construction, the forward-rate or swap-rate distribution implied by selected market option prices can be fitted exactly at calibration dates.

But “exact fit” has boundaries:

• quotes are discrete in strike;
• a smile interpolation/extrapolation is required;
• market bid–ask spreads mean there is not one unique true surface;
• numerical quadrature/interpolation introduces error;
• only the chosen calibration instruments are fitted exactly.

The calibration surface becomes part of the model.

13. Inputs and outputs

Inputs can include:

• initial discount curve;
• swap schedules and annuities;
• swaption/caplet strikes and market prices or implied vols;
• smile interpolation/extrapolation rule;
• state-process mean reversion/volatility parameters;
• state correlation structure in multifactor versions;
• numeraire choice;
• calibration dates;
• state grid/quadrature nodes;
• boundary extrapolation rules.

Outputs can include:

• state-to-swap-rate/forward-rate mappings;
• bond-price functions P(t,T;X);
• reproduction errors for calibration instruments;
• implied terminal rate distributions;
• Bermudan/exotic option values;
• exercise boundaries;
• state sensitivities;
• grid/quadrature convergence diagnostics;
• smile-dynamics stress results.

14. The hidden modelling choice is dynamics between calibrated marginals

Suppose the model exactly matches every one-year and two-year swaption smile.

Many different stochastic dynamics can share those same marginal distributions.

The Markov state process and mapping choices determine:

• how today’s smile evolves tomorrow;
• correlation between different swap rates;
• joint movement across exercise dates;
• Bermudan continuation values;
• hedging behaviour.

Therefore exact vanilla calibration does not identify exotic dynamics uniquely.

15. Mean reversion becomes a dynamics parameter

In a one-factor Gaussian Markov driver, a parameter such as mean reversion controls how strongly state values at different dates remain correlated.

Changing mean reversion can leave each calibration-date smile nearly unchanged after remapping, yet materially change the joint dynamics between dates.

Bermudan prices can therefore be sensitive to mean reversion even when European swaption fit remains exact.

Diagnostic: stress mean reversion across a plausible range and report the resulting exotic-value sensitivity.

16. One factor creates strong dependence restrictions

If every market rate is a function of one scalar state, their contemporaneous movements are heavily constrained.

A one-factor model can be extremely efficient but may not represent:

• independent level/slope/curvature moves;
• basis movements;
• decorrelation between distant swap rates;
• multi-curve risks.

Falsifier: compare model-implied co-movement with historical/market correlation evidence and hedge performance. Large systematic residuals signal that one factor is too restrictive.

17. Evidence polarity

Evidence for confidence includes:

• calibration instruments reprice within tolerance;
• state-to-rate mappings remain monotone;
• discount bonds remain positive and internally consistent;
• today’s discount curve is reproduced;
• digital probabilities are nonnegative and monotone;
• quadrature/grid refinement changes prices negligibly;
• Bermudan values agree with independent benchmarks in simpler cases;
• exotic prices are stable under reasonable smile interpolation choices.

Evidence against confidence includes:

• non-monotone mapping;
• negative or inconsistent bond prices;
• large calibration residuals;
• strong exotic-price sensitivity to arbitrary smile wings;
• large mean-reversion/model-dynamics sensitivity;
• poor hedge performance despite exact vanilla fit;
• grid/quadrature instability;
• inconsistent results across numeraire implementations.

18. Counterexample: exact smile fit, wrong Bermudan value

Two Markov-functional calibrations can reproduce the same set of European swaption prices but use different mean-reversion parameters.

Their marginal swap-rate distributions at calibration dates can be the same while the joint transition dynamics differ.

A Bermudan option—whose value depends on conditional future exercise opportunities—can therefore have different prices.

Falsifier: stress dynamics parameters that do not materially affect vanilla calibration. If exotic values move strongly, report that model risk explicitly.

19. Counterexample: smile extrapolation drives the result

Market quotes cover only a finite strike range. The Markov-functional mapping requires a distribution across a wider state region.

If deep-wing smile extrapolation is arbitrary, the state-to-rate map in extreme regions is also arbitrary.

Falsifier: compare several static-arbitrage-consistent wing assumptions. If the exotic value or risk changes materially, the tail is not market identified.

20. Counterexample: noisy digital extraction

A smile fitted directly through noisy quotes can produce unstable strike derivatives. The implied digital probability curve can become non-monotone.

A non-monotone CDF cannot support a valid quantile mapping.

Falsifier: check digital probabilities and implied density before building the state mapping. Smoothness without static-arbitrage consistency is insufficient.

21. Counterexample: mapping loses monotonicity

The one-to-one quantile mapping assumes swap rate increases consistently with the Markov state.

If numerical interpolation creates local reversals, the inverse mapping can become multi-valued.

Falsifier: test discrete derivatives of S(X) over the full state grid. A sign change violates the monotone-mapping assumption.

22. Counterexample: one factor cannot hedge curve shape

A Bermudan swaption can respond to multiple sections of the yield curve. A one-state model may reproduce value but produce unrealistic hedge ratios across tenors.

Falsifier: compare model hedge P&L under historical level, slope and curvature shocks. Persistent unexplained hedge error indicates missing factors.

23. Markov-functional versus Hull–White

Hull–White algorithms specify the short-rate dynamics directly and derive rates/bonds from that diffusion.

Markov-functional models can use a similar Gaussian state but obtain flexibility by mapping the state nonlinearly into market rates to fit a chosen smile distribution.

The two approaches can share state-process intuition while assigning flexibility to different parts of the model.

24. Markov-functional versus HJM

HJM algorithms specify the dynamics of the entire forward curve and enforce the no-arbitrage drift restriction.

Markov-functional models instead compress pricing into a low-dimensional Markov state and calibrated bond/rate functions.

25. Markov-functional versus the LIBOR Market Model

LIBOR Market Model algorithms model many forward rates directly, with drift coupling and a high-dimensional correlation structure.

Markov-functional modelling trades some direct forward-rate dynamical specification for a lower-dimensional state and exact-fit functional mapping.

26. Markov-functional versus Jamshidian decomposition

Jamshidian decomposition exploits one-factor monotonicity to decompose certain swaptions into bond options under suitable models.

Markov-functional models also exploit low-dimensional monotonic structure, but their defining job is the calibrated functional mapping and backward construction of bond functions.

27. Exercise pricing

Once the state functions are built, a Bermudan option can be priced by backward induction on the state grid.

At exercise date t_j:

V_j(X) = max(ExerciseValue_j(X), ContinuationValue_j(X)).

The continuation value is a conditional expectation under the chosen measure.

Because the state is one- or low-dimensional, this backward induction can be much faster than full high-dimensional forward-rate simulation.

28. Numerical diagnostics

• curve reconstruction: reproduce today’s bond prices;
• vanilla repricing: reproduce calibration swaptions/caplets;
• digital monotonicity: tail probabilities must decrease with strike;
• mapping monotonicity: X→S mapping must preserve the chosen order;
• state-grid convergence: refine X grid;
• quadrature convergence: increase conditional-integration nodes;
• tail truncation: widen state boundaries;
• mean-reversion stress: test exotic sensitivity;
• smile-wing stress: vary extrapolation;
• benchmark: compare with simpler analytical/tree models where possible.

29. Weak links

• noisy or arbitrage-inconsistent smile interpolation;
• incorrect numeraire/measure conversion;
• non-monotone state mapping;
• state grid too narrow;
• quadrature insufficient;
• mean reversion treated as unimportant because vanilla fit is exact;
• one-factor dynamics used for multi-factor hedge jobs;
• tail extrapolation hidden;
• discount curve or annuity conventions inconsistent;
• multi-curve basis ignored when material.

30. What would falsify confidence?

Confidence should be withdrawn if the model cannot reproduce calibration instruments; if the state-to-rate mapping loses monotonicity; if bond prices become inconsistent; if Bermudan/exotic values are unstable under grid refinement; if plausible mean-reversion or smile-wing changes produce large unexplained value swings; or if hedging evidence shows that the low-dimensional dynamics omit important curve movements.

31. Verification and update triggers

Preserve discount curves, smile quotes, interpolation model, numeraire, state-process parameters, state grid, quadrature rules, mapping tables, calibration residuals and exotic benchmark results.

Revalidate when:

• market smile conventions change;
• discount/projection curve architecture changes;
• mean-reversion calibration changes;
• smile wings become materially different;
• negative-rate/shift conventions change;
• new Bermudan/exotic products are added;
• hedge performance deteriorates;
• state-grid or numerical libraries change.

32. Primary and high-quality references

• Phil Hunt, Joanne Kennedy and Antoon Pelsser, Markov-Functional Interest Rate Models, original working-paper version.
• Hunt, Kennedy and Pelsser, Markov-Functional Interest Rate Models, SSRN version associated with the published framework.
• Joanne Kennedy, Phil Hunt and Antoon Pelsser, Markov-Functional Interest Rate Models, Finance and Stochastics, 2000.
• Federal Reserve, Supervisory Guidance on Model Risk Management, for validation and ongoing monitoring principles.

Educational boundary: A Markov-functional model can fit selected vanilla option distributions extremely well while still making strong assumptions about joint dynamics between calibration dates. Exact calibration is not proof of correct exotic-option behaviour.