CMT common mistakes
CMT · Understand the traps and how to avoid them.
Fading a strong trend with an oscillator: shorting a stock because RSI(14) prints above 70 while price is above a rising 200-day MA and making higher highs.
Overbought/oversold levels are the easiest rule to memorize, so candidates apply them as standalone signals and treat 'extreme' readings as reversal calls without first classifying the market regime; strong trends keep oscillators pinned for weeks.
What to do instead
Classify regime before choosing a tool: price vs. 200-day MA, MA slope, and ADX. In a confirmed uptrend, use oscillator dips (not peaks) as entries and only take countertrend oscillator signals after a trend break is confirmed by price structure.
Calling a chart pattern complete before its confirmation trigger: labelling a head-and-shoulders 'triggered' at the second shoulder, or on an intraday poke below the neckline that closes back above it.
Pattern names invite premature labelling, and candidates answer from the picture's shape rather than from the pattern's completion criteria, which the exam tests explicitly.
What to do instead
For every pattern, write down three things before answering: confirmation (a close beyond the neckline/trendline plus volume expansion), invalidation (a close back inside the pattern), and the measured-move target. Only act when confirmation is met.
Mixing timeframes: treating a daily 50/200 moving-average cross as a short-term trading signal, or treating a small weekly penetration of support as a breakdown.
Multi-timeframe reasoning is cognitively expensive under time pressure, so candidates collapse everything onto the chart in front of them and forget that the same signal carries very different weight on a weekly versus a 5-minute chart.
What to do instead
State the timeframe explicitly for each signal you use, work top-down (weekly context, then daily setup, then intraday trigger), and only act when the higher timeframe does not contradict; remember MA crossovers are lagging by construction.
Misclassifying biases in vignettes: tagging anchoring, confirmation, or conservatism as emotional biases because they 'feel' emotional.
Candidates memorize bias names and examples but not the underlying taxonomy, and the two categories overlap in ordinary language.
What to do instead
Sort by mechanism, not by feel: cognitive errors stem from faulty reasoning or information processing (conservatism, confirmation, representativeness, anchoring, hindsight, framing, mental accounting) and are correctable with better information; emotional biases stem from feelings or impulses (loss aversion, overconfidence, self-control, regret aversion, status quo, endowment) and are harder to correct. Rehearse by sorting 20 short vignettes into the two buckets.
Reversing prospect theory's risk attitudes: stating that investors become risk-averse when facing losses, or that the value function is convex in gains.
The shape of the value function is easy to recall visually but the direction of curvature is easy to flip under exam pressure, especially when the question frames the choice in terms of a sure loss.
What to do instead
Anchor on the four facts: concave in gains (risk-averse), convex in losses (risk-seeking), loss limb steeper than gain limb (loss aversion, roughly 2:1), and outcomes evaluated relative to a reference point. Test yourself with the classic framed gamble (sure loss vs. 50/50 larger loss) until the prediction is automatic.
Predicting the wrong behaviour from a bias: expecting a client with the disposition effect to cut losing positions, or expecting an overconfident trader to reduce position size after a loss.
Candidates learn the bias label without learning its observable footprint in trading records, then answer with the 'rational' behaviour instead of the biased one.
What to do instead
Memorize each bias as a measurable behaviour: disposition effect = high proportion of realized gains and low proportion of realized losses plus high turnover; overconfidence = high turnover, under-diversification, and narrow forecast intervals; regret aversion = excess cash and reluctance to act. Match the data signature, not the label.
Treating VaR as the maximum possible loss, or quoting a VaR figure without its confidence level and horizon.
VaR is a single familiar number, and '95% VaR' sounds like a ceiling rather than a quantile; the tail beyond the quantile is invisible unless you look for it.
What to do instead
Always quote VaR fully: '1-day 95% VaR of $2m' means a 1-in-20 chance of losing $2m or more, not a cap. Pair every VaR answer with expected shortfall/CVaR and a stress scenario, and remember VaR is not sub-additive so portfolio VaR can exceed the sum of component VaRs.
Sizing positions from stop distance in points or currency rather than risk per share, and ignoring correlation: 'risk 2% per trade' applied to eight correlated technology longs.
The 2%-per-trade rule is easy to recall but the denominator (risk per share) is easy to forget, and position-level rules feel sufficient without a portfolio-level overlay.
What to do instead
Compute risk per share as |entry − stop|, then size = (account equity × risk fraction) ÷ risk per share, adjusting for gap risk and slippage. Then aggregate: cap total portfolio heat and haircut correlated positions (e.g., treat eight 0.8-correlated longs as roughly one to two independent bets) before adding a new trade.
Annualizing or averaging returns incorrectly: multiplying a monthly return by 12, averaging multi-period returns arithmetically, or mixing arithmetic and geometric means in the same comparison.
Simple multiplication and averaging are the default mental operations, and the arithmetic/geometric distinction is taught once and then blurred by calculator shortcuts.
What to do instead
Link returns geometrically for multi-period performance, annualize with (1 + R)^(252/n) − 1 for daily data or (1 + R)^(12/n) − 1 for monthly, and state which mean you are using: geometric for realized past performance, arithmetic for forward-looking expected returns and for inputs to portfolio optimization.
Presenting an overfitted backtest as evidence: adding parameters until the equity curve is smooth, optimizing and testing on the same sample, or using a current-index constituent list that excludes delisted names.
A smooth equity curve is psychologically persuasive and the exam rewards recognizing statistical artifacts, but in practice the feedback loop of tweaking until it looks good is fast and invisible unless you deliberately separate samples.
What to do instead
Use a strict hold-out plus walk-forward validation, freeze the parameter count in advance, report a deflated or multiple-testing-adjusted Sharpe ratio, and rebuild the universe point-in-time to remove survivorship and look-ahead bias before trusting any result.