Reader question: A Monte Carlo path is observed at 100 and then 102. If a knock-out barrier is at 95, can we safely conclude the asset never touched 95 between those two simulated dates?

No. The endpoints do not reveal the entire continuous path. A diffusion can cross a barrier between grid dates and return before the next simulated observation. Standard time-stepped Monte Carlo therefore tends to miss some barrier events when a contract is continuously monitored.

A Brownian bridge supplies the missing conditional probability. Given two simulated endpoints, it asks: what is the probability that the continuous diffusion crossed the barrier somewhere between them? That probability can be used either to simulate an intra-step crossing or to weight the payoff by the probability of survival.

This article owns the intra-step barrier-monitoring correction problem: simulated diffusion endpoints + barrier + local variance + monitoring convention → crossing probabilities, survival weights, corrected barrier events and diagnostics for monitoring bias.

It does not own generic Sobol path construction, the pricing model’s volatility calibration, or discrete-monitoring contract definitions. Those are separate roles. The purpose here is to distinguish path-simulation resolution from the legal monitoring rule and to show how a one-line probability can repair a large numerical bias.

This is public mathematical education about derivative-pricing algorithms. It is not investment advice or a recommendation to buy or sell barrier products.

1. Why endpoint monitoring is biased for continuous barriers

Suppose a down-and-out option disappears if the asset ever touches barrier H before maturity.

A naive Monte Carlo algorithm simulates:

S₀, S₁, …, S_N

and declares the option knocked out only if some sampled node satisfies:

S_i ≤ H.

But a continuously monitored contract is knocked out when:

min_0≤t≤TS_t ≤ H.

Those are different events.

With a coarse grid, the node test misses crossings. For a knock-out option, missed crossings usually bias value upward because too many simulated paths survive.

2. Move to log price under geometric Brownian motion

Under Black–Scholes dynamics:

dS/S = (r−q)dt + σdW.

Let:

X = ln S.

Then:

dX = (r−q−½σ²)dt + σdW.

For a down barrier H, define:

h = ln H.

Conditional on two endpoints X_t=x and X_t+Δ=y, the path between them is a Brownian bridge with volatility scale σ.

The remarkable result is that the conditional barrier-crossing probability has a simple form.

3. Lower-barrier crossing probability

If both endpoints lie above a lower log barrier h:

x > h, y > h,

then the conditional probability that the bridge touches or crosses the barrier during the interval is:

p_hit = exp[−2(x−h)(y−h)/(σ²Δ)].

If either endpoint is already at or below the barrier, the hit probability is one.

Several features are immediately sensible:

if either endpoint approaches the barrier, p_hit approaches one;
if both endpoints are far above the barrier, the probability becomes tiny;
if Δ becomes small while distances stay fixed, the crossing probability shrinks;
if σ rises, hidden intra-step crossing becomes more likely.

4. Upper barriers are symmetric

For an upper log barrier h with both endpoints below it:

x < h, y < h,

the conditional crossing probability is:

p_hit = exp[−2(h−x)(h−y)/(σ²Δ)].

The same structure appears: multiply the endpoint distances from the barrier, divide by interval variance, and place the result inside an exponential.

5. Why the drift disappears from the conditional formula

The log-price process has drift, yet the bridge crossing formula above does not contain r, q or the log drift explicitly.

Conditioning on both endpoints removes the deterministic linear drift component from the bridge fluctuation. The residual uncertainty about the path shape depends on the diffusion variance over the interval.

This is a powerful general lesson: conditional simulation can simplify a problem by conditioning away variables that would matter unconditionally.

6. Method A: simulate the hidden crossing with a uniform draw

For each pair of endpoints that remain on the safe side of the barrier:

compute p_hit;
draw U ~ Uniform(0,1);
if U < p_hit, mark the barrier as crossed;
otherwise continue the path.

This produces an explicit random barrier-event indicator consistent with the conditional crossing probability under the model assumptions.

The extra Bernoulli draw adds variance, though. If we care only about expected payoff, there is often a more efficient route.

7. Method B: survival-probability weighting

For a knock-out option, the conditional probability of surviving one interval is:

p_survive,i = 1 − p_hit,i.

Conditional on the simulated endpoints, the path’s continuous-monitoring survival probability across all intervals is:

P_survive = ∏_i(1−p_hit,i),

under the standard conditional decomposition interval by interval.

Instead of drawing Bernoulli barrier events, multiply the terminal payoff by this survival weight.

This is a conditional-Monte-Carlo or Rao–Blackwell-style improvement: integrate out some random barrier-event uncertainty analytically rather than simulate it.

For many knock-out payoffs this reduces variance and makes the price smoother with respect to model inputs.

8. Knock-in options can use in/out relationships when valid

For standard contracts with matching terms and no complications, an in/out parity can relate knock-in, knock-out and vanilla options.

For example, under compatible definitions:

vanilla = knock-in + knock-out.

This gives a valuable verification invariant. Price the knock-out with Brownian-bridge monitoring, price the knock-in independently or by parity, and check reconciliation.

Rebates, monitoring windows and contract-specific clauses can modify simple parity, so the legal payoff specification must be checked before using it.

9. Continuous monitoring and discrete monitoring are different products

A Brownian-bridge correction is appropriate when the contract is continuously monitored but the simulation is discrete.

If the legal contract checks the barrier only at the end of each day, month or fixing date, a crossing between those dates should normally not trigger the contractual barrier.

Using a continuous Brownian bridge for a legally discrete barrier would “correct” the simulation toward the wrong product.

This is one of the most important authority boundaries in numerical derivatives work:

first identify the contractual monitoring convention; then choose the numerical algorithm.

10. Broadie–Glasserman–Kou continuity correction addresses the opposite mapping

Sometimes a closed-form continuous-barrier formula is available but the contract is discretely monitored.

Broadie, Glasserman and Kou showed that discrete barrier prices can often be approximated accurately by shifting the barrier used in the continuous formula.

The magnitude of the multiplicative adjustment is:

exp(βσ√Δt),

where:

β ≈ 0.5826.

The barrier is shifted away from the current underlying when mapping a discretely monitored contract to a continuous-barrier approximation: upward for an upper barrier and downward for a lower barrier.

This makes sense because a discrete monitor is less likely to observe a hit than a continuous monitor.

The continuity correction and Brownian-bridge intra-step correction solve related but different numerical problems. They should not be mixed without understanding which contract each approximation targets.

11. Inputs and outputs

Inputs can include:

spot and barrier levels;
upper/lower barrier direction;
knock-in or knock-out payoff;
continuous or discrete contractual monitoring rule;
simulation time grid;
local interval volatility;
model type: GBM, local volatility, stochastic volatility, jump model;
rebate and timing rules;
number of paths and random-number configuration.

Outputs can include:

per-interval hit probabilities;
simulated barrier-event flags;
path survival weights;
barrier-option price;
Monte Carlo standard error;
monitoring-grid convergence table;
difference versus naive node monitoring;
difference versus analytic or PDE benchmark.

12. A worked conditional crossing example

Suppose the asset moves from 100 to 102 over an interval, with a lower barrier at 95 and constant volatility σ.

In log space:

x = ln 100, y = ln 102, h = ln 95.

The crossing probability is:

exp[−2(x−h)(y−h)/(σ²Δ)].

If Δ is one day and volatility is modest, the probability may be small but not zero. If the endpoints sit at 96 and 97 instead, the endpoint distances from the barrier shrink dramatically and the hidden crossing probability can become substantial.

The algorithm therefore reacts to the geometry that node monitoring ignores.

13. Near-barrier numerical stability deserves attention

When an endpoint is extremely close to the barrier, p_hit is near one. Floating-point roundoff can create values slightly outside [0,1] if formulas are rearranged poorly.

Robust implementations:

short-circuit immediately if an endpoint breaches the barrier;
compute in log space where appropriate;
clamp only tiny numerical overshoots caused by rounding, not model errors;
test intervals with endpoints exactly at the barrier;
avoid taking logs of survival probabilities after they have numerically become zero without handling the case explicitly.

14. Products with many monitoring intervals can underflow survival products

If:

P_survive = ∏(1−p_hit,i)

contains thousands of factors, direct multiplication can underflow.

A numerically stable implementation accumulates:

log P_survive = Σ log(1−p_hit,i)

and exponentiates only when needed.

This is a useful example of how probability theory and numerical analysis meet: a mathematically correct product can still fail in floating-point arithmetic.

15. Local volatility: freeze the interval cautiously

Under a local-volatility model, σ depends on time and asset level. The simple constant-σ bridge formula is no longer globally exact.

A practical approximation can use a locally frozen or averaged variance over each short interval. The shorter the interval and the smoother the local-vol surface, the more plausible that approximation becomes.

But “Brownian bridge correction applied” is not enough information. Validation must specify which local variance was used and how error changes as the time grid is refined.

16. Heston and stochastic volatility add another hidden path

Under Heston, both asset price and variance evolve between the endpoints. Conditional on only asset endpoints, the integrated variance over the interval is not fixed.

A bridge crossing formula that simply inserts one endpoint variance is approximate.

More sophisticated Heston barrier algorithms combine variance simulation, integrated-variance approximations and conditional sampling. The Brownian-bridge idea still helps, but its exact GBM formula should not be presented as exact under stochastic variance.

17. Jump models can invalidate the continuous-diffusion bridge

A jump can cross a barrier instantly. Brownian bridge formulas describe continuous diffusion between endpoints.

In a jump-diffusion model, the algorithm must account for jump times and jump sizes as well as continuous crossing between jumps.

In a pure-jump infinite-activity model, the hitting problem can be more fundamentally different.

Falsifier: if jumps are part of the model, show explicitly where jump-driven barrier crossings enter. A plain Brownian bridge alone is incomplete.

18. Evidence polarity

Evidence for confidence includes strong agreement with analytic continuously monitored Black–Scholes barrier prices, much smaller time-grid bias than naive node monitoring, stable values as the grid is refined, lower Monte Carlo variance from survival weighting than Bernoulli crossing when applicable, correct in/out parity, and smooth behaviour as the barrier moves far away.

Evidence against confidence includes prices that change materially when the simulation grid is refined, Brownian-bridge results that disagree with analytic GBM benchmarks, a bridge applied to a legally discrete barrier, use of constant-volatility crossing formulas inside a stochastic-volatility or jump model without validation, negative/greater-than-one hit probabilities, or Greeks that remain highly unstable despite apparently converged prices.

19. Counterexample: daily nodes miss an overnight excursion

A path closes Monday at 100 and Tuesday at 101. A down barrier is 98. Node monitoring says “safe.”

The diffusion could have fallen to 97.5 overnight and recovered by Tuesday’s node.

Falsifier: compare the node-monitoring price with Brownian-bridge-adjusted Monte Carlo and with an analytic continuous-barrier benchmark. The node bias should become visible, especially when volatility is high or the barrier is close.

20. Counterexample: over-correcting a discretely monitored contract

A monthly barrier is legally checked only on twelve observation dates. A developer applies Brownian-bridge continuous hitting probabilities between those dates.

The code now prices a more restrictive continuous-monitoring feature rather than the contractual monthly feature.

Falsifier: reconstruct the payoff directly from the term sheet. If between-date crossings have no contractual effect, the continuous bridge event should not be triggered.

21. Counterexample: price converges but delta does not

Barrier payoffs are discontinuous around hitting events. A small change in spot can change which paths cross the barrier, causing noisy finite-difference Greeks.

Survival weighting often smooths the expectation, but Greek estimators still need independent validation.

Falsifier: compare bump-and-revalue deltas across bump sizes and common-random-number seeds, and benchmark against PDE/analytic Greeks where available.

22. Counterexample: a far barrier makes corrections irrelevant

If a down barrier sits extremely far below every material path, node monitoring and Brownian-bridge monitoring can agree within Monte Carlo noise.

This is not evidence that node monitoring is generally safe. It means the test case is insensitive.

Falsifier: validate with barriers both far from and close to spot, across several volatilities and maturities.

23. Counterexample: a barrier near spot exposes step-size dependence

When the barrier begins close to spot, hidden crossing probabilities can be large in the first few intervals. A coarse-grid naive simulation can be severely biased.

Falsifier: halve the time step repeatedly. The naive estimator should move toward the corrected benchmark; a bridge-corrected estimator should show much smaller monitoring bias under GBM.

24. Alternatives

Analytic barrier formulas are ideal benchmarks for standard continuously monitored Black–Scholes barriers.

Finite-difference PDE methods place a knock-out barrier directly into the spatial domain as an absorbing/rebate boundary.

Binomial/trinomial trees can represent barrier events but may converge irregularly if the barrier is poorly aligned with nodes.

Continuity correction approximates discrete-monitoring prices from continuous formulas without Monte Carlo.

Conditional sampling can restructure the simulation to reduce barrier-event variance further, especially in quasi-Monte-Carlo settings.

25. Verification ladder

Contract test: confirm whether monitoring is continuous, discrete or windowed.
Endpoint breach test: hit probability must be one if a node is already beyond the barrier.
Probability bounds: every computed crossing probability must lie in [0,1].
Analytic GBM benchmark: compare continuously monitored vanilla barrier prices with closed-form results.
Grid test: compare daily, half-daily, hourly or progressively finer grids.
Naive-versus-bridge test: quantify missed-crossing bias.
Bernoulli-versus-survival test: verify equal means within sampling error and compare variance.
In/out parity: reconcile where contract definitions permit.
Far-barrier limit: recover the corresponding vanilla value as the barrier becomes irrelevant.
Model extension test: separately validate local-vol, Heston or jump approximations.

26. Connections to the surrounding Bukit Timah Tutor knowledge estate

The generic simulation engine is covered by Monte Carlo pricing algorithms. Brownian-bridge barrier correction is a path-dependent refinement inside that machinery.

Sobol quasi-Monte Carlo algorithms also use the phrase “Brownian bridge,” but often for a different job: constructing path coordinates so important variance is concentrated in early low-discrepancy dimensions. This article uses Brownian bridges to estimate barrier crossing between known endpoints. The two ideas can coexist but should not be conflated.

Finite-difference option-pricing algorithms provide an independent deterministic route and a powerful benchmark because a knock-out barrier becomes a boundary condition.

The underlying diffusion baseline is developed in Black–Scholes option-pricing algorithms.

Stochastic-volatility path generation connects to Heston calibration and model-specific simulation methods.

The full lane is indexed at Finance & Banking Algorithms | Applied Mathematics in Real Financial Systems.

27. What would falsify confidence?

Confidence should be withdrawn if a supposedly continuous-monitoring Monte Carlo price fails an analytic Black–Scholes barrier benchmark; if values retain material grid dependence after bridge correction; if survival weighting and Bernoulli crossing disagree beyond Monte Carlo error; if the algorithm silently changes the legal monitoring convention; or if a GBM crossing formula is used in a stochastic-volatility/jump model without an error study.

28. Verification and update triggers

Preserve the monitoring convention, barrier level, rebate rule, model, volatility input, path grid, bridge formula, crossing implementation, random-number configuration and analytic/PDE benchmark results for each product class.

Revalidate after changes in contractual monitoring frequency, local/stochastic volatility dynamics, jump assumptions, time-grid construction, Greek methodology, barrier interpolation or QMC path construction. Trigger review when barriers move close to spot, volatility rises materially, monitoring windows become irregular, or a product introduces multiple moving barriers.

29. Primary and high-quality references

Mark Broadie, Paul Glasserman and Steven Kou, A Continuity Correction for Discrete Barrier Options, Mathematical Finance, 1997.
Mark Broadie, Paul Glasserman and S. G. Kou, Connecting Discrete and Continuous Path-Dependent Options, Finance and Stochastics, 1999.
Paul Glasserman, Monte Carlo Methods in Financial Engineering, Springer, 2003.
Nico Achtsis, Ronald Cools and Dirk Nuyens, Conditional Sampling for Barrier Option Pricing under the LT Method, SIAM Journal on Financial Mathematics.
Bruno Casella and Gareth O. Roberts, Monte Carlo Simulation of Killed Diffusions with a Single Barrier, work using crossing probabilities and Brownian-bridge/Rao–Blackwell ideas.

Educational boundary: Brownian-bridge correction repairs a monitoring discretisation problem under specified diffusion assumptions. It does not repair a wrong payoff definition, a wrong volatility model or jump dynamics that the bridge formula does not contain.

Small Group Tutorials

How Brownian-Bridge Barrier Algorithms Detect Missed Option Crossings: Conditional Hitting Probabilities, Survival Weighting, Continuity Corrections and Monitoring Bias