On this page
FRM® Part II portfolio construction asks how an active return forecast becomes an implementable set of weights. A screen, stratification, linear programming and quadratic programming impose progressively more explicit controls, but each gives up something: flexibility, data simplicity or robustness to estimation error. The right examination answer starts by identifying which risk and cost terms the method actually handles.
Start with the decision, not the algorithm
Suppose a manager expects one security to outperform its benchmark weight. A larger position raises expected active return, but it can also increase exposure to a sector, a common factor or an illiquid name. Trading into the position has a cost. The portfolio decision therefore concerns the entire set of positions and their interactions, not just the rank of that security. GARP places risk management and investment management within Part II, which emphasises applying risk tools to decisions rather than recalling isolated definitions. Its official study materials and learning objectives set out that curriculum context.
A useful way to read the construction problem is to separate three inputs. First, the alpha forecast says where expected return may come from. Second, the risk model says how active positions move together and how far the portfolio may depart from its benchmark. Third, transaction costs say what it takes to get from current holdings to target holdings. A constraint, such as a sector limit or maximum position, describes the feasible set. An objective chooses the preferred portfolio within that feasible set. Confusing these roles makes an optimiser seem more magical than it is.
What each construction method actually controls
A screen orders securities by forecast alpha and admits those above a cut-off or within a top group. It is transparent and easy to explain. It need not produce a well diversified active portfolio, because many high-ranked securities may share a driver. A buy, hold and sell band can moderate turnover: a name already held need not be sold merely because its rank slipped slightly. Even then, the method does not directly price the covariance of selected holdings or the aggregate cost of trading. It is a selection rule, not a full risk objective.
Stratification starts by dividing the benchmark into cells, often by sector and size. The manager selects favoured securities within each cell while keeping the cell weights close to benchmark weights. This prevents a good stock ranking in one sector from silently becoming a large sector bet. The price is lost freedom: a strong forecast for a whole sector cannot be expressed if that sector is fixed at its benchmark weight. It also only controls the attributes used to make the cells. Two securities in the same cell may still have very different exposures to an unmodelled factor.
Linear programming represents several requirements as linear relationships in the portfolio weights. It can require benchmark-like exposure to named characteristics, impose position bounds and include a linear approximation to trading costs. Its attraction is that the constraints can be stated precisely and solved together without constructing an unwieldy grid of cross-classified cells. The answer remains only as good as the characteristics selected. A portfolio can meet every listed limit while carrying risk in a factor nobody listed. Linear exposure matching also says nothing by itself about how many names the solution holds.
Quadratic programming brings the covariance matrix into the objective. A portfolio variance is quadratic in its weights because the risk contribution from one active position depends on the other positions. This makes the method better suited to balancing forecast alpha against joint active risk, alongside costs and explicit limits. It is more informative than simply penalising each position separately. Yet the mathematically optimal weights are optimal for the estimated inputs. If covariance estimates or alpha forecasts are noisy, the program may concentrate in an apparent opportunity that disappears when the inputs are revised.
An illustrative constraint changes the answer
Consider a deliberately simplified benchmark with two sectors, each with half its weight. A manager has an illustrative active-return forecast of four percentage points for security A in the first sector and two percentage points for security B in the second. The manager can add a total active overweight of ten percentage points, financed by underweights elsewhere. If the only objective were to maximise forecast alpha and neither risk nor costs mattered, putting all ten points into A would add an estimated 0.4 percentage point of active return. That is a calculation about a hypothetical forecast, not a market prediction.
Now suppose the mandate permits at most six percentage points of active overweight in either sector. The unconstrained ten-point allocation to A is infeasible. A simple feasible allocation puts six points in A and four in B. The forecast active contribution is 0.06 times 0.04 plus 0.04 times 0.02, or 0.32 percentage point. The constraint therefore reduces the forecast contribution by 0.08 point in this artificial example. It may still improve the portfolio decision if it removes an unwanted sector concentration. A candidate should be able to explain that sacrifice without asserting that the constrained portfolio is always better ex post.
Add an illustrative trading cost of 0.05 percentage point for establishing the two positions and the estimated net contribution becomes 0.27 point. If a risk penalty were also applied, it would be deducted from that figure. The example shows why a ranking alone cannot finish the task: feasibility is checked before comparison, and the preferred feasible portfolio depends on costs and risk as well as return. A quadratic approach would need variances and covariance for A and B, plus the risk aversion parameter. Without those inputs it cannot honestly produce a numerical optimum.
How to diagnose a proposed solution
First ask what the active weights are measured against. A position can be large in absolute terms and small relative to the benchmark, or the reverse. Active risk concerns departures from the benchmark, so a constraint on total weight does not automatically limit active sector exposure. Then ask whether the constraint is hard or soft. A hard cap rules out a portfolio; a penalty merely makes violating a preference costly. An examination description that says an exposure must remain inside a band implies a feasibility condition, not just a lower score.
Next identify the risk information actually available. If only broad category labels are known, stratification may be a defensible simple control. If several measured factor exposures are known, linear programming can match them. If a usable covariance matrix is available and joint risk matters, quadratic programming can represent it. This is an information distinction, not a universal ranking of quality. The official GARP risk principles for asset managers stress defining a portfolio objective and monitoring estimated investment risk. A construction method does not remove the need for that judgement.
Check turnover separately. A portfolio that looks attractive from target weights can lose much of its expected advantage if it requires repeated trading. A rating band, explicit cost term or minimum-trade rule addresses different aspects of this problem. Do not assume that a method which matches benchmark exposures has also priced trading. Conversely, including a cost estimate does not guarantee that the estimate captures market impact under stressed liquidity. Costs are conditional estimates, and constraints can become binding as prices, forecasts and holdings change.
The exam mistake: treating detail as certainty
The most common conceptual mistake is to call the most sophisticated optimiser the safest answer in every setting. The quadratic method uses more of the available relationship between holdings, but it also demands more estimates. Its output can be highly sensitive to small errors when many securities have similar forecast returns. A practical check is to vary the inputs and ask whether the recommendation changes radically. A stable, slightly less ambitious allocation can be easier to implement and monitor than one built on fragile precision. That is a decision principle, not a claim that simple screens always outperform.
Another mistake is to treat category neutrality as comprehensive neutrality. Matching sector and size cells controls those labels, but it does not ensure matching duration, liquidity, momentum or any other exposure not represented in the cells. Equally, a screen that preserves the strongest alphas can accidentally combine correlated positions. In a short exam prompt, underline the words that specify the benchmark, permitted deviations, cost of trading and available covariance information. The omitted item often identifies the method’s weakness.
For revision, practise describing the trade-off in one sentence before doing any algebra: what signal is kept, what risk is restrained, and what cost or model error remains. Then test a tiny invented portfolio like the one above. The FRM program page gives the broader programme context, while Omni’s methodology explains how its study approach is organised. Keep the underlying question in view: what portfolio can actually be held, and why does the chosen construction rule make that portfolio preferable under the stated assumptions?
Sources
- FRM Study Materials | GARP · garp.org · Retrieved
- FRM Learning Objectives | GARP · garp.org · Retrieved
- Risk Principles for Asset Managers | GARP · garp.org · Retrieved



