Numerical methods turn a model definition into a number that a desk, risk process, or client can use. Interviews therefore probe two kinds of understanding: how the algorithm works and how you know the output is trustworthy.
For every method, prepare its assumptions, convergence behavior, failure modes, stopping rule, computational cost, and validation strategy. A number of decimal places is not an accuracy argument unless you can account for model, input, discretization, sampling, and floating-point error.
What this guide helps you do
- Decompose numerical error and choose defensible tolerances.
- Compare root finding, interpolation, PDE, Monte Carlo, and optimization methods.
- Design convergence and benchmark tests for pricing routines.
- Discuss production performance without weakening correctness.
1. Begin with error and conditioning
Separate model error, input error, discretization error, iterative-solver error, sampling error, and floating-point error. Tightening a solver tolerance does not repair a misspecified model or noisy market input. Report the component relevant to the decision.
Conditioning describes how sensitive the true problem is to input changes; stability describes how the algorithm propagates errors. A stable implementation cannot make an ill-conditioned calibration informative, but it can avoid adding unnecessary error.
- Use absolute tolerances near zero and relative tolerances at scale.
- Check dimensions, signs, monotonicity, bounds, and conservation identities.
- Compare refinements to estimate convergence rather than trusting one grid.
- Retain an independent reference or high-accuracy benchmark for regressions.
2. Root finding and interpolation
Bisection is robust when a continuous function is bracketed and changes sign; Newton can converge rapidly near a simple root but depends on derivatives and a reasonable start; secant methods trade derivative requirements for superlinear convergence. Hybrid methods often combine a safe bracket with faster local steps.
Interpolation choices should respect financial shape. Linear interpolation is transparent but nonsmooth. Cubic splines are smoother but can overshoot. Monotone or shape-preserving methods help when discount factors, survival probabilities, or total variance must obey constraints.
| Method | Strength | Failure or control |
|---|---|---|
| Bisection | Guaranteed bracket contraction | Requires a valid sign-changing bracket; linear convergence |
| Newton | Fast local convergence | Derivative, start, flat slope, and leaving the domain |
| Secant | No analytic derivative required | Less robust and may leave a safe region |
| Hybrid | Robust bracket with faster accepted steps | More logic; still needs domain and stopping checks |
3. Finite-difference PDE methods
A PDE solver requires a spatial domain, time grid, boundary and terminal conditions, discretization scheme, linear solver, and interpolation back to the requested state. Explicit schemes are simple but conditionally stable; implicit schemes are robust but require solving systems; Crank-Nicolson is higher order in smooth settings but can oscillate around nonsmooth payoffs without damping.
Test domain truncation, grid refinement, boundary sensitivity, positivity or monotonicity where relevant, and convergence to an analytical price. For American exercise, explain the free-boundary or complementarity problem and the selected method.
4. Monte Carlo and variance reduction
Monte Carlo is flexible in high dimension and for path dependence, but its standard error usually falls only as one over the square root of the number of independent samples. Report an estimator with a standard error or confidence interval, not just a price.
Antithetic variables, control variates, stratification, importance sampling, quasi-random sequences, and conditional expectation can reduce variance. Each method relies on structure and should be validated for bias and effectiveness on the actual payoff.
- Separate path discretization bias from estimator sampling error.
- Construct and test correlations, including near-singular cases.
- Manage random streams and seeds for reproducibility and parallel safety.
- Use analytical moments and prices as unit and convergence checks.
5. Calibration, optimization, and production quality
Calibration is an inverse problem. Define instruments, quote conventions, weights, constraints, parameter transformation, objective, and acceptance criteria. Multiple parameter sets may fit similarly, so inspect identifiability, stability, residual structure, and sensitivity to starting values.
In production, cache only with correct invalidation, vectorize without creating uncontrolled temporary memory, parallelize independent work, and preserve auditability. Monitor convergence failures, fallback use, stale inputs, parameter jumps, runtime, and differences from a benchmark.
- Start with synthetic parameters to verify recovery under controlled noise.
- Compare global exploration with local refinement when objectives are non-convex.
- Use parameter bounds or transformations that reflect the model domain.
- Distinguish a low objective value from economically acceptable residuals.
Practise aloud
Interview drills with answer direction
Question 1
When would you use bisection instead of Newton?
Answer direction: Use bisection when a valid bracket and robustness matter more than speed, especially with weak derivatives or poor initial guesses. A hybrid can preserve the bracket while accepting faster steps.
Question 2
How do you show a PDE price is converged?
Answer direction: Refine time and space grids separately, vary boundaries, compare with an analytical or independent benchmark, inspect expected order, and test price and Greeks rather than one output only.
Question 3
Why is Monte Carlo attractive in high dimensions?
Answer direction: Its basic sampling convergence rate does not directly deteriorate with dimension like a tensor grid, and path-dependent states are easy to simulate. Variance and per-path cost can still be large.
Question 4
What makes calibration ill-conditioned?
Answer direction: Parameters may have similar effects on available instruments, data may be noisy or sparse, and the objective may be flat in some directions. Diagnose with sensitivities, repeated starts, profile objectives, or singular values.
Question 5
How should a numerical tolerance be chosen?
Answer direction: Relate it to scale, conditioning, downstream materiality, discretization and input error, and achievable floating precision. Verify that tightening the tolerance no longer changes the decision-relevant output.
Turn reading into practice
A focused study plan
- Module 1
Accuracy
Create an error budget and practise conditioning, stability, tolerance, and floating-point checks.
- Module 2
One-dimensional tools
Implement bracketed roots, Newton, and shape-aware interpolation with tests.
- Module 3
Pricing engines
Build a PDE or tree solver and a Monte Carlo estimator against the same analytical benchmark.
- Module 4
Calibration
Fit synthetic and market-like data, analyze residuals and parameter stability, and document fallbacks.
Self-review
Frequent mistakes to catch early
- Equating solver convergence with a correct or well-conditioned answer.
- Reporting a Monte Carlo estimate without uncertainty.
- Testing only one grid, starting point, or market state.
- Optimizing runtime while silently changing discretization or tolerance.
Continue with structured practice
Relevant Desk2Quant resources
Numerical Methods for Quants: The Master Field Manual
Work through desk-oriented root finding, interpolation, PDE, Monte Carlo, optimization, and calibration implementations.
Explore this resourceGreeks, Vols, Yield Curves, Numerical Methods, Monte Carlo & XVA Guide
Connect numerical method choices to volatility, curves, Greeks, simulation, and XVA calculations.
Explore this resourceUltimate Industry-Grade Quant Project Pack
Choose further projects to demonstrate convergence, testing, performance, and model-risk judgment.
Explore this resourceKeep building
Related quant finance guides
Common questions
Frequently asked questions
Which numerical methods should a quant know?
Core topics include root finding, interpolation, linear algebra, optimization, finite differences, trees, Monte Carlo, random-number methods, SDE discretization, calibration, error analysis, and numerical testing.
What numerical methods appear in quant interviews?
Interviewers commonly ask about bisection and Newton, interpolation, Monte Carlo error and variance reduction, PDE schemes, calibration, floating point, convergence, and validation against benchmarks.
How do I choose between PDE and Monte Carlo pricing?
Use product features, state dimension, path dependence, exercise, required Greeks, accuracy, and runtime. PDE methods are strong in low-dimensional Markov settings; Monte Carlo is flexible for high-dimensional and path-dependent problems.
What makes numerical code production-ready?
It has explicit contracts, stable algorithms, convergence evidence, analytical or independent benchmarks, boundary tests, justified tolerances, failure handling, versioned inputs, observability, and repeatable performance tests.