Multi-Objective Optimization and Pareto Fronts

Most real optimization problems have more than one goal, and the goals conflict. A portfolio manager wants high return and low risk; a trading system wants high fill rate and low market impact. Optimizing either goal alone is easy, but pursuing both at once means accepting a trade-off, and there's no single number that says which trade-off is "correct" — that depends on how much risk (or latency, or cost) the decision-maker is willing to tolerate.

A solution is called Pareto-efficient if no other feasible solution improves one objective without making another one worse. The set of all such solutions is the Pareto front: a curve or surface of options, each one optimal for some particular weighting of the objectives, none of them dominating any other.

The classic finance example is the mean-variance efficient frontier: for a fixed level of risk, one portfolio weighting maximizes expected return, and no other portfolio at that same risk level does better. Move along the curve and both the risk and the best-achievable return change together; nothing on the curve is more "optimal" than anything else on it without specifying how much risk you're willing to bear.

A portfolio manager comparing two candidate weightings — one at 8% expected return with 12% volatility, another at 10% return with 16% volatility — cannot say which is "better" from the numbers alone. Both can sit exactly on the efficient frontier; the only way to pick between them is to decide, separately, how much extra volatility is worth two extra points of expected return. That decision is a risk preference, supplied by the investor, not something the optimization itself can resolve.