Full Replication, Sampling and Optimization

Full replication: when owning everything is easy

For the S&P 500 or the EURO STOXX 50, full replication is trivial: a few hundred hyper-liquid names, minimal turnover (Chapter 5's self-maintenance), negligible tracking friction. Even MSCI World's ~1,300 names are comfortably replicable at scale, and the largest world funds hold essentially all of them. Full replication's virtues: zero model risk (nothing to get wrong), maximal transparency, and tracking limited only by fees, taxes and the small frictions of Chapter 35's waterfall.

Its cost appears at the tails. An index like MSCI ACWI IMI (~8,500 names) or FTSE Global All Cap (~9,500) includes thousands of micro-positions: a 0.001% weight in a thin frontier-market small cap, where the round-trip spread might be 2%, custody in the local market costs real money, and the position's contribution to index behavior is statistically indistinguishable from zero. Buying it all is possible; it's just paying tolls for cargo that doesn't matter.

Sampling: the craft of the representative subset

Stratified sampling, the workhorse in bonds: divide the index into cells (by sector, rating, maturity bucket, currency) and hold enough liquid bonds in each cell to match the cell's weight, duration and yield. A global aggregate index may contain nearly 30,000 bonds, many of which haven't traded in weeks; the flagship funds tracking it hold a fraction of that count, chosen so the characteristics match: duration to a hundredth, spread duration, rating mix, sector weights. The bond market's structure makes this not merely acceptable but necessary: full replication of a broad bond index is a fiction no one has ever run.

Optimization, the equity version: a risk model (factor exposures, sector/country weights, covariances) selects the subset minimizing expected tracking error subject to holding fewer names and avoiding the costliest tails. An ACWI IMI fund might hold ~3,000–4,000 of 8,500 names: all the large and mid caps outright, then a curated small-cap sample carrying the statistical signature of the thousands omitted.

Where sampling can bite: the honest weakness

The omitted tail is chosen to be statistically irrelevant in normal correlations. Sampling's weak moments are when the tail stops behaving:

• Dispersion events: a small-cap rally concentrated in exactly the omitted names, a single frontier market repricing violently, a credit-market episode where the unheld off-the-run bonds gap differently from the held on-the-runs. Result: a quarter of unusual tracking deviation, positive or negative.
• Index events at the tail: a tiny omitted name becoming an acquisition target with a 60% pop contributes a sliver of index return the fund didn't earn.
• Model risk proper: an optimizer calibrated on yesterday's covariances meeting a regime change. Rare, real, and the reason TE (Chapter 35's consistency statistic) earns its narrow keep here: a sampled fund's TE tells you how loosely the copy is allowed to fit.

The empirical comfort: for the flagship sampled funds, realized multi-year TDs sit within a few basis points of fully replicated peers, and frequently better, because the avoided trading costs exceed the sampling noise. The craft, at scale, has been winning its bet for two decades. The vigilance: sampled funds on exotic indices from small providers are where the bet occasionally shows its teeth, and where your Chapter 35 history-reading matters most.

How to see what your fund does

1. The factsheet's replication line (Chapter 24): "optimized/sampled" stated plainly.
2. Holdings count vs index count, both on the factsheet: 1,380 of 1,395 is effectively full; 2,900 of 8,500 is a serious sample; 480 of 2,600 on an EM index is a bold one, and its TD history had better be good.
3. Multi-year TD and its year-to-year spread (Chapter 35): the craft's report card. A sampled fund with tight, consistent TD has earned trust; one with good averages but wild individual years is showing you its model risk.
4. Stress-episode checks: how did the fund track through March 2020, or 2022's bond repricing? Annual reports and TD histories answer it: the sample's behavior when correlations broke is the closest thing to a crash test you can read.

A worked example

Ravi compares two funds on FTSE Global All Cap (~9,500 names):

• Fund X: 9,200 holdings, TER 0.24%, 5-yr TD −0.21%, spread −0.25 to −0.18.
• Fund Y: 3,700 holdings, TER 0.22%, 5-yr TD −0.16%, spread −0.22 to −0.11, including 2020.

Fund Y holds 40% of the census and tracks better: the tail's trading costs, avoided, outweighed the sampling noise, exactly the craft thesis. Ravi's residual question is the right one: what does Y's TE and stress history look like? Both benign, per the factsheet and 2020's annual report. He takes Y, and records the meta-lesson in his notes: completeness is an aesthetic; characteristics-matching plus a verified report card is an engineering standard, and this book sides with engineering.