schur

← NCO bridge demos paper (PDF) · verification script

The best dial in closed form, and the sign that decides whether small noise moves it

Identical groups reduce the whole problem to one number, and two conditions from the paper become exact: the optimal dial under a two-point estimate is (1 + 2(k−2)c) / (2[1 + (k−3)c]), and under small noise the dial leaves full coupling only if a computable coefficient is positive.

The setting

There are k groups of assets. Each group contains two: a representative, which the paper calls the group's knot, and an independent asset. The representatives of different groups are correlated with each other, every pair at the same correlation c. The independent assets are uncorrelated with everything. All representatives have variance 1; all independent assets have variance 1/δ, so a large δ means a safe independent asset.

Because every group looks the same, a sensible portfolio treats them the same. It puts some total weight x on the representatives and the rest, 1 − x, on the independent assets, spread evenly. The whole allocation problem is therefore one number, x. The lowest-variance choice is x⋆ = 1 / (1 + δL), where L = 1 + (k − 1)c measures how much the representatives move together, and the variance paid for missing it grows with the square of the miss.

Two methods, and the bridge between them

Nested clustered optimization (NCO) first builds a portfolio inside each group using only that group's own covariance, then allocates across the groups. Inside a group it cannot see that the representative is correlated with the other groups' representatives. Global minimum variance uses the whole covariance at once. The bridge is a family of portfolios indexed by a dial γ between 0 and 1: at γ = 0 it is NCO, at γ = 1 it is global minimum variance, and in between each group is told a fraction γ of what the other representatives explain about its own.

The correlation is estimated

The true correlation c is unknown. The portfolio is built from an estimate z, and then it has to live under the true c. The score is therefore the expected out-of-sample variance F(γ): build the bridge portfolio from the estimate, compute its variance under the truth, and average over what the estimate might have been. Two kinds of estimate are considered.

2 0.25
1.00
0.050
The covariances. Left, the true correlation: the representatives K1…Kk are correlated at c, the independent assets (m) are not correlated with anything. Middle and right, the two estimates the portfolio might be built from. Only the block of representatives differs between them.
The representative weight. The weight x that the bridge puts on the representatives, against the dial γ, on each of the two estimates. The grey line is the ideal weight x⋆ under the true correlation. The bridge aims for that line while seeing only the estimate.
The out-of-sample variance. The left end is NCO, the right end is global minimum variance. If the lowest point is strictly inside, then trusting the estimated correlation only partly beats both trusting it fully and ignoring it. With the two-point estimate the paper gives the best dial in closed form, γ⋆ = (1 + 2(k−2)c) / (2[1 + (k−3)c]), drawn in red; the numerical minimum is drawn in green and sits on top of it. A green marker off the red line would refute the closed form.
With small symmetric noise the red line is a prediction from an expansion in τ, not an exact optimum: it is drawn only when it lands on the bridge, and the costs in the readout are always measured against the numerical minimum. The τ slider is capped so that both estimates remain covariances, since an equicorrelated estimate is one only for −1/(k−1) < z < 1. At the default k = 2, c = ¼, δ = 1 the best dial is exactly 2/3. The reason is visible in the first chart: on the estimate 0 the bridge cannot move, and on the estimate 2c the dial 2/3 lands the weight exactly on x⋆.

When a little damping helps, and when it does not

For small noise the paper shows the best dial is 1 − (G′(1)/V₀″(1))τ²: a fixed coefficient times the noise squared, below 1. If the coefficient is positive, full coupling is beaten by a slightly smaller γ. If it is negative, the formula points past 1 and full coupling is best. The chart plots that coefficient against δ for the current k and c.

The sign of the coefficient. Above the zero line, small noise pushes the best dial inside; below it, full coupling stays best. For k = 10 and c = ¼ the paper gives the coefficient exactly, (8 − δ)/(1 + 13δ/4), so the crossing is at δ = 8. Set k = 10 and δ = 12: small noise then leaves the best dial at 1, yet switching to the two-point estimate moves it to 10/11. Whether damping helps depends on how noisy the estimate is, not on the method alone.

Every curve is the paper's closed-form scalar expressions evaluated live; nothing is simulated. The same numbers are checked in exact rational arithmetic by the verification script.