CQF common mistakes
CQF · Understand the traps and how to avoid them.
In Monte Carlo or binomial pricing, the candidate writes down an answer that ignores the discounting convention implied by the question (e.g. multiplies an expectation by e^{-rT} when the payoff was already expressed in forward or T-forward measure terms, or discounts from the wrong date), producing a price that is off by a clean factor like e^{-rT} or e^{+rT}.
Under exam time pressure the mechanics of building the tree or simulating paths absorb attention, and the measure/discounting step is treated as an afterthought. Candidates also conflate the risk-neutral drift with the discount factor and apply both twice, or once too few.
What to do instead
Before computing, write one line stating the numeraire and the discount date, e.g. 'price = N_0 * E^{N}[payoff / N_T]'. After computing, run a sanity check: a zero-coupon bond payoff of 1 must price to the discount factor, and a forward contract with zero strike must price to the discounted forward. If either fails, the measure or discounting is wrong.
When estimating volatility or a covariance matrix from a short return sample, the candidate plugs the raw sample estimate straight into a portfolio optimisation or a VaR calculation without shrinking it, without accounting for estimation error, and without noticing that the resulting weights are extreme (large long/short positions, negative variance estimates).
The sample covariance is the textbook estimator and candidates treat it as the true parameter. They do not connect the noise in a T-observation, N-asset covariance matrix to the instability of the inverse, especially when N is comparable to T.
What to do instead
State the ratio N/T explicitly and apply a shrinkage estimator (e.g. Ledoit-Wolf toward a diagonal or constant-correlation target) or a factor model. Check that the optimised weights are economically plausible; if they are not, report the shrunk solution and say why. For VaR, use a t-distribution or filtered historical simulation rather than a raw normal fit to a short sample.
In a regression-based or ML forecasting question, the candidate fits the model on the full sample, reports in-sample R-squared or accuracy as the performance measure, and does not use a walk-forward or purged split, so overlapping labels leak future information into training.
Standard cross-validation is the default habit from general ML, and financial labels (e.g. forward returns over a horizon) overlap, so random k-fold splits leak across the boundary. Candidates also equate fit with predictive skill.
What to do instead
Always use time-ordered splits with a purge/embargo gap equal to at least the label horizon, and report out-of-sample metrics only. Compare against a naive benchmark (random walk for returns, persistence for volatility). Treat any in-sample R-squared above roughly 0.05 for return prediction as a red flag, not a success.
The candidate answers a question about a stochastic process by quoting the distribution of the terminal value but then applies Itô's lemma incorrectly, dropping the second-order term (the 1/2 sigma^2 S^2 d^2V/dS^2 term) when deriving a dynamics or a PDE, or applying the correction with the wrong sign.
The ordinary chain rule is deeply practised and the Itô correction is easy to forget under pressure, particularly when the function is simple like ln S or S^2 and the answer 'looks right'. Sign errors arise from differentiating a decreasing function.
What to do instead
Write the Itô expansion in full every time: dV = V_t dt + V_S dS + 1/2 V_SS d. Then substitute d = sigma^2 S^2 dt before simplifying. Verify with a known case: for V = ln S the drift must be (mu - sigma^2/2), and for V = S^2 the drift must be (2mu + sigma^2)S^2.
For a Greeks or hedging question, the candidate reports delta, gamma, vega as if they were constant and sizes a hedge on that basis, ignoring that the hedge must be rebalanced and that gamma/vega change with the underlying and with time; or the candidate confuses the sign conventions (e.g. states that a long call has negative gamma, or that vega is negative for a long option).
Greeks are memorised as static formulas rather than as sensitivities evaluated at a point. Sign conventions are learned by rote and collapse under pressure, and the connection between gamma and the cost of rebalancing (the gamma-theta relationship) is often not internalised.
What to do instead
For every Greek, state the sign for long call, long put, short call, short put in a small table before answering, and derive from the payoff shape if unsure. When asked about hedging, explicitly mention rebalancing frequency and the gamma P&L: 0.5 * Gamma * (dS)^2 versus Theta * dt. Check that a delta-hedged long option position has positive gamma and negative theta.
In a fixed income question, the candidate computes duration and convexity but applies them to a yield change measured in the wrong units (basis points versus decimal), or applies modified duration to a parallel shift while the question specifies a curve twist, or ignores the difference between Macaulay and modified duration.
The formulas are simple and the unit conversion is the step most often skipped. Candidates also default to parallel-shift thinking because it is the textbook case, even when the question describes a steepening or a key-rate shift.
What to do instead
Write the yield change as a decimal (25 bp = 0.0025) before multiplying, and state which duration you are using: Macaulay / (1 + y/m) for modified. If the question describes a non-parallel shift, use key-rate durations or reprice under the specified curve scenario rather than a single duration number.
When asked to evaluate a strategy or model, the candidate reports a Sharpe ratio or hit rate computed on a sample without adjusting for multiple testing, without a transaction-cost assumption, and without a benchmark, and concludes the strategy is profitable.
Performance metrics are taught as single-number summaries, and the deflated-Sharpe and multiple-testing corrections are less practised. Candidates also forget that a backtest with daily rebalancing can be entirely consumed by costs.
What to do instead
Always state the number of strategies or parameters tried and apply a deflation (e.g. deflated Sharpe ratio or a Bonferroni-style haircut). Include a realistic cost assumption and report gross and net. Compare against the relevant benchmark and report the sample length, since a high Sharpe on 12 months is not evidence.
In a risk management question, the candidate computes VaR but does not state the confidence level, horizon, and method, and then treats VaR as a coherent or subadditive measure, or reports expected shortfall as if it were bounded above by VaR.
VaR is introduced first and its limitations (non-subadditivity, silence about the tail beyond the quantile) are covered later and are easy to skip. Candidates also mix up the direction of the inequality for ES.
What to do instead
State 'VaR at the X% level over an H-day horizon, computed by historical simulation / parametric / Monte Carlo'. Remember VaR is not subadditive in general and ES is; ES is at least as large as VaR at the same level. If a question asks about tail risk, default to ES and say why.
On a machine learning question, the candidate selects a model by accuracy on an imbalanced classification problem (e.g. predicting defaults or regime shifts where the positive class is a few percent), and reports a high accuracy that simply reflects the majority class.
Accuracy is the default metric in general ML practice, and the imbalance in financial labels is severe and often not stated explicitly in the question. Candidates do not compute the base rate before evaluating.
What to do instead
Compute the base rate first (predicting the majority class gives that accuracy). Then use precision, recall, F1, or AUC, and for trading-relevant questions use a cost-weighted or profit-based metric. State the confusion matrix or at least the base rate alongside any accuracy figure.
When a question involves calibration or parameter estimation (e.g. fitting a volatility surface, a Hull-White mean reversion, or a GARCH model), the candidate reports the fitted parameters without checking identifiability, stability, or whether the calibration reproduces market quotes, and does not mention the objective function or weighting used.
Calibration is often treated as a black-box step, and candidates do not distinguish an ill-posed inverse problem from a well-posed one. In GARCH, near-unit-root persistence and small samples produce unstable estimates that look plausible.
What to do instead
State the objective (e.g. weighted squared errors in implied vol), the weights, and the resulting fit error. Check parameter plausibility against theory (GARCH alpha+beta < 1 and not implausibly close to 1; mean reversion positive). For surface fits, report the max repricing error in vol points and note any arbitrage violations (negative implied variance, calendar arbitrage).