On this page
A binomial tree price becomes unreliable when the result changes materially with the number of steps, the volatility input, the discount curve, or the assumed shape of volatility through time. For a CAIA® Level II candidate, the useful answer is to diagnose which assumption moves the valuation, then state what evidence or sensitivity check would reduce that uncertainty. A neatly computed backward-induction price does not establish that the model represents the security well.
What the tree actually answers
A tree maps possible future values of an underlying asset into payoffs and works backwards to a value today. The arithmetic is conditional: given the up and down moves, a probability consistent with the model's discounting convention, and the contractual payoff at every node, what price follows? CAIA Association's curriculum overview places modelling in its methods area, while its Level II study guide describes tree-based valuation of convertibles and callable bonds in an earlier curriculum. The dated guide establishes the modelling context; candidates should use their current assigned curriculum for the live exam scope.
The distinction between a calculation mistake and model risk matters. A candidate may apply backward induction perfectly yet begin with a volatility estimate that is inappropriate for the payoff. Equally, a sound estimate of volatility cannot rescue a tree with so few time steps that exercise opportunities are represented at the wrong dates. Model risk is the uncertainty introduced by the structure and inputs of the valuation method itself. It calls for a reasoned range and a diagnosis, not another pass through the same calculator sequence.
Step size is a modelling choice
A binomial tree replaces a continuous range of possible paths with a finite grid. The shorter each step, the more finely the grid can represent terminal payoffs and exercise opportunities. A coarse grid can put an important strike or call threshold between its reachable nodes. A one-step model may therefore give an appreciably different option value from a many-step model even when both use the same headline volatility and rate. Increasing the step count is a way to test numerical stability, but the output need not move smoothly in one direction at every increase.
Imagine an illustrative option with a payoff only when a stock finishes above a stated strike. With a very coarse tree, a small change in step count can move one terminal node from below the strike to above it. That changes the set of positive payoffs abruptly. With additional steps, the terminal distribution is represented more finely, and the price may settle around a stable range. The point of the exercise is not to memorise a particular number of steps. It is to compare successive resolutions and decide whether the remaining price change is material to the decision being examined.
A convergence check needs a fixed set of economic assumptions. Keep the maturity, rate, volatility estimate, payoff and exercise rule the same while changing only the number of steps. Record the resulting value and the difference from the previous grid. If the differences are still consequential, the step count has not earned confidence. If they become small, discretisation has been reduced, but the other model choices remain open. A candidate who calls a stable grid an accurate price has skipped that second part of the test.
A worked sensitivity example
Consider an invented, one-period, European-style claim used only to show the logic. Its underlying starts at 100, the up node is 120, the down node is 80, and the claim pays 20 in the up state and zero in the down state. Assume a one-period risk-free gross return of 1.05. Under this simple no-arbitrage setup, the risk-neutral up weight is the gross return less the down factor, divided by the up factor less the down factor: (1.05 − 0.80) divided by (1.20 − 0.80), or 0.625. The model value is the weighted payoff of 12.50 divided by 1.05, approximately 11.90. These are invented teaching numbers, not an observed market price.
Now change only the assumed spread of the two moves: use an up factor of 1.30 and a down factor of 0.70, with a claim payoff of 30 in the up state and zero below the same strike of 100. The risk-neutral up weight becomes (1.05 − 0.70) divided by (1.30 − 0.70), about 0.5833. The discounted weighted payoff is about 16.67. The value rose by roughly 4.76 even though the underlying's starting level, strike, maturity and gross risk-free return were unchanged. This is a deliberately extreme sensitivity demonstration; it shows that uncertainty about the dispersion of possible outcomes can matter more than precise arithmetic at the original nodes.
The example isolates parameter sensitivity, not a real volatility estimate. In an actual tree, the up and down factors may be derived from an assumed volatility and time interval. The rate may influence the model's risk-neutral weighting and discounting. These are related inputs, so a disciplined comparison changes one assumption at a time and checks that the resulting weights remain economically valid. Never treat the risk-neutral weight as the actual frequency with which the stock will rise. It is a pricing device under the specified model, not a forecast of investor beliefs.
Volatility can be wrong even when the grid converges
Volatility is estimated from some evidence and mapped into the tree. Historical variation may not describe the future, and a volatility inferred from traded options may differ by maturity or strike. A single constant input simplifies all the nodes into one dispersion rule. That simplification can be especially consequential for a payoff that responds mostly to unusually high stock prices, unusually low interest rates, or an early call decision. The region where the option is sensitive can matter more than an average volatility across all paths.
For a convertible, the value of conversion depends on the stock path, while the bond component and issuer-related terms introduce additional assumptions. For a callable bond, a fall in rates can make the issuer's call option important, so a tree that misrepresents the distribution of low-rate paths can misstate the embedded option. The earlier CAIA Level II guide explicitly identifies both security types as tree applications. These examples explain why it is worth asking where the payoff changes sharply, rather than asking only whether the tree was built correctly.
Rate and payoff conventions need their own check
The discount rate is not a harmless label. The rate and compounding convention used to construct a risk-neutral tree must fit the time step and the claim being valued. A mismatch between an annual rate and a shorter step can change the probability calculation or the discount factor. For a fixed-income security, the relevant curve and spread assumptions can be more complex than the single rate used in a classroom example. State the convention before doing the calculation; otherwise two apparently different model prices may simply reflect inconsistent input definitions.
The payoff rule must also be implemented at the right node and time. A callable bond's issuer decision, a holder's conversion choice, a coupon payment and a maturity redemption are different cash-flow events. If a call is allowed only on specified dates, inserting exercise at every tree node gives the issuer an option it does not have. A finer grid cannot fix a wrong contract. The useful diagnostic is to write down the contractual decision at each allowed time, verify the node payoff, and only then work backwards. The CAIA programme page provides the broader study context; the methodology page explains how to organise practice around decisions.
A compact exam decision process
When an exam prompt reports a surprising tree value, first separate what was held fixed from what changed. If the step count alone changed, investigate discretisation and convergence. If the spread of node outcomes changed, investigate the volatility assumption. If the discount curve or compounding convention changed, inspect both risk-neutral weights and discount factors. If the legal exercise window changed, inspect the payoff specification. Name the likely direction only when the payoff and other assumptions justify it; option sensitivity is not a licence to make a universal claim about every structured security.
For revision, make a small comparison table in your notes with columns for assumption, affected tree component, check and remaining limitation. Step count affects resolution; compare increasingly fine grids. Volatility affects node dispersion; compare plausible input ranges. Rates affect weights and discounting; reconcile the curve and time convention. Exercise terms affect node payoffs; read the contract. Then explain in one sentence what the test did not resolve. This is the difference between reporting a single apparently exact output and evaluating whether the output deserves reliance.
The common mistake is to conclude that more steps cure all model risk. More steps can reduce numerical approximation error within a chosen model. They do not verify a constant-volatility assumption, repair a stale volatility estimate, choose a suitable discount curve or correct the exercise terms. For the assigned CAIA concept, the strongest answer connects the price change to its source and proposes a check that directly addresses that source. That conclusion is transferable from a simple option to the more involved tree-based securities described in the CAIA materials.
Sources
- Explore the Curriculum | CAIA Association · caia.org · Retrieved
- CAIA Level II Study Guide March 2023 | CAIA Association · caia.org · Retrieved



