AQIF common mistakes
AQIF · Understand the traps and how to avoid them.
Annualising a volatility estimate by multiplying the periodic standard deviation by the number of periods (e.g., monthly sigma x 12) rather than by the square root of the number of periods.
Candidates memorise 'annualise by multiplying by 12' from return calculations and carry the same scaling rule over to a dispersion measure, ignoring that variance scales linearly with time while standard deviation scales with its square root.
What to do instead
State the scaling rule explicitly before computing: variance scales with t, so sigma_annual = sigma_period x sqrt(periods per year). Sanity-check the answer against a plausible range (e.g., a 4% monthly sigma should give roughly 14%, not 48%).
Interpreting a p-value as the probability that the null hypothesis is true, e.g., writing 'there is a 5% chance the strategy has no skill' after obtaining p = 0.05.
The p-value is the most-quoted statistic in the curriculum and its definition is easy to recite but easy to invert under time pressure, especially when the question is phrased in decision language.
What to do instead
Rehearse the conditional direction: p = P(data at least as extreme as observed | H0 true). It says nothing about P(H0 | data). If a question asks for the probability the null is true, the answer requires a prior and Bayes' rule, not the p-value.
Estimating the price change of a bond from a large yield shift using modified duration alone, e.g., applying -D_mod x delta_y to a 250 bp move and reporting the result as the expected price change.
Duration is the first term taught and the easiest formula to reach for; the convexity correction is treated as a refinement rather than a required second-order term, and its sign and factor of one-half are often misremembered.
What to do instead
Always add the second-order term: delta_P/P is approximately -D_mod x delta_y + 0.5 x C x (delta_y)^2. Use duration-only only when explicitly told the shift is small, and check that the convexity term is positive for a plain vanilla bond in both rate directions.
Fitting a scaler, imputer, or feature-selection step on the full dataset before splitting into train and test, then reporting cross-validated performance from that pre-processed matrix.
The convenient workflow (load, clean, scale, split) matches how libraries are demoed, and the leak is invisible in the code because the transform is a single familiar line rather than an obvious use of the target.
What to do instead
Put every fitted transform inside a pipeline and let cross-validation refit it on each training fold only. As a check, ask whether any statistic used to build a feature was computed using rows the model will later be tested on; if yes, refit within the fold.
Reporting overall classification accuracy as evidence of a good model on a dataset where the positive class is rare, e.g., claiming 98% accuracy for a default model when 2% of firms default.
Accuracy is the default output of most libraries and feels like a single, comparable number, so it gets quoted without comparison to the trivial always-predict-majority baseline.
What to do instead
Report the confusion matrix alongside precision, recall, and a threshold-free measure such as AUC, and always state the no-skill baseline (the majority-class rate). If accuracy is close to that baseline, the model has added nothing.
Describing a 95% one-day VaR of $10m as 'the maximum we can lose' or 'the loss we expect on a bad day'.
VaR is introduced as a single number summarising downside, and the quantile definition is compressed into a headline that sounds like a worst case; the 5% exceedance region gets lost.
What to do instead
Translate every VaR figure into the full statement: losses are expected to exceed $10m on about 5% of days (roughly one day in 20), and VaR says nothing about how large those exceedances are. Pair it with expected shortfall when the question concerns tail severity.
Running an unconstrained mean-variance optimisation on raw historical sample mean returns and sample covariance, then presenting the resulting concentrated corner solution as the optimal portfolio.
The optimisation is treated as a mechanical calculation, so the estimation error in the inputs, which dominates the output, is never considered; the mathematics rewards extreme weights on assets with the highest estimated return-to-risk ratio.
What to do instead
Treat the optimiser as an error amplifier: apply shrinkage or Black-Litterman to expected returns, impose position and turnover constraints, and test sensitivity of weights to small perturbations of the inputs. If weights swing wildly, the inputs, not the optimiser, are the problem.
Sizing positions by signal strength or by equal capital allocation across assets with very different volatilities, so that the highest-volatility names dominate the portfolio's risk.
The signal is the intellectually interesting part of the strategy, and position sizing is treated as an afterthought, so capital weights are confused with risk weights.
What to do instead
Convert every position to an ex-ante risk contribution: scale each position by the inverse of its volatility (or by a target risk budget) and verify that realised risk contributions, not dollar weights, match the intended allocation.
Backtesting a strategy on the index's current constituent list, or on the set of funds and stocks that still exist today, and reporting the resulting return stream as achievable.
Free and convenient data sources provide only the latest snapshot of a universe, and survivorship bias is invisible in the resulting equity curve because the delisted and failed names simply never appear.
What to do instead
Use point-in-time constituent and security-master data, or reconstruct the universe as of each rebalance date including names later removed. State explicitly in the methodology how delistings and returns of failed firms were handled.
Selecting the best-performing configuration from dozens or hundreds of parameter combinations and reporting that configuration's Sharpe ratio or hit rate without any adjustment for the number of trials.
The search is treated as legitimate research rather than a statistical test, and the maximum of many noisy estimates looks impressive precisely because it is a maximum.
What to do instead
Count and disclose the number of configurations tested, report the full distribution of results rather than the winner, and apply a multiple-testing adjustment (deflated Sharpe ratio, Bonferroni-style haircut) or confirm the result on a genuinely untouched holdout period.
Continuing to direct order flow to a broker after receiving hospitality, or receiving confidential information about an issuer and trading on it after labelling the situation a 'mosaic'.
Each individual step feels normal and incremental, and the conflict or the materiality of the information is rationalised after the fact rather than assessed before acting.
What to do instead
Apply the test before acting: would an independent observer see this as a conflict, and is the information material and non-public with the source under a duty of confidentiality? Disclose gifts and hospitality to compliance and pre-clear personal and related-party trades; if either test is uncertain, escalate rather than trade.