Nested Clustered Optimization

Also seen as: NCO

Why split the optimisation into groups?

Because estimation errors spread through a single large optimisation. When several instruments are almost identical, a minimum-risk optimiser over all of them at once can turn tiny differences in their measured correlations into large offsetting bets. Inside a small group the problem is well behaved, and the groups are then compared as a handful of synthetic instruments — so a measurement error inside one group cannot move the weights inside another.

where: $w_{i \mid k}$ is instrument i's weight inside its group k, $W_k$ is group k's weight across the groups, and $w_i$ is the instrument's final share of the capital. Both levels aim at the lowest-risk mix, the "Min volatility" end of the efficient frontier.

The number of groups is chosen from the data at each rebalance, not set by you.

When does Nested Clustered Optimization fit, and when not?

It fits a universe of many similar instruments — a basket of stocks that fall into sectors, say — where the groups are clear and a single optimisation would be most fragile. It fits less well in two cases:

• A universe with no clear groups. The two-stage answer then drifts further from what a single lowest-risk optimisation would give.
• A book built on a hedge. Two instruments that move in opposite directions are treated as maximally different and land in different groups, so the method cannot fully use the hedge between them.

And like any lowest-risk optimisation, the across-group step can give a whole high-risk group a weight close to zero.

How does Fincanva handle it?

• Nested Clustered Optimization is offered inside a single strategy only, across its instruments; a Combined does not offer it.
• It has no settings of its own. It reads the calculation window (In-sample, 12 months by default) and the Risk estimation choice — see Risk estimation.
• Weights are never negative.
• With most risk estimates — including the recommended Ledoit-Wolf · constant correlation — every instrument must have moved in price at some point inside the window: an instrument whose price stayed flat for the whole window, such as a suspended listing, stops the backtest rather than receiving a weight. Two estimates are the exception; Risk estimation names them.
• Weights are recomputed at every rebalance from the window ending on that date.

Which plan includes Nested Clustered Optimization?

Included from Advanced upwards. See what each plan includes.

What does it look like in practice?

A strategy holds six instruments that fall into two groups of three: three bank stocks and three utility stocks. Inside the bank group, the lowest-risk mix is 50% / 30% / 20%; inside the utility group it is 40% / 40% / 20%. Treated as two synthetic instruments, the groups are then mixed for the lowest risk, which here comes out at 30% banks and 70% utilities.

The first bank stock's final weight is 50% × 30% = 15%; the first utility stock's is 40% × 70% = 28%. Every one of the six weights is built the same way, and together they add up to 100%.