Smart Beta and the Factor Zoo

You’ve spent five lessons building the factor machine: CAPM cracked, size and value bolted on, momentum and profitability and investment piled on top, regressions to estimate loadings, and a hard look at how much “alpha” is really just hidden factor beta. Now comes the reckoning. The same research engine that discovered a handful of real, durable factors has also published hundreds of them — and turned factor exposure into a product you can buy in one click. This final lesson is the skeptic’s capstone: how factors got packaged into smart-beta ETFs, why the academic literature is drowning in a factor zoo of 300-plus published anomalies, how reliably those anomalies decay once they’re discovered and traded, and why timing factors is, for almost everyone, a losing game. The throughline: the durable factors are the old, boring, theoretically grounded ones, not the newest entry in the zoo.

Smart beta: beta you choose

Analogy. Think of the salad bar of investing. A plain index fund hands you the pre-made house salad — the whole cap-weighted market, take it or leave it. A traditional active manager is a chef who custom-builds your plate and charges you for the talent. Smart beta is the salad bar in between: a menu of factor tilts — value, momentum, quality, low-volatility — that you assemble à la carte, paying for the recipe, not the chef.

Definition. A smart-beta (or strategic-beta / factor) ETF is a rules-based, transparent fund that deliberately tilts its holdings toward one or more documented factors instead of weighting by market capitalization. It sits squarely between passive indexing and active management: like passive, it’s mechanical, cheap, and rules-driven; like active, it intentionally deviates from the market to chase a premium. The key reframing: it is beta you choose, not alpha you’re sold. You’re not buying skill; you’re buying systematic exposure to a risk factor whose premium was already in the academic literature.

The common menu items map straight onto the factors from earlier lessons:

Why “beta you choose” is the right mental model. If a value ETF beats the market over a decade, the factor framework says: of course it might — you deliberately loaded on HML, and HML has historically carried a premium. That outperformance is not alpha; a factor regression would absorb almost all of it into the value loading and leave a residual near zero. You chose a tilt and got paid (or not) for the tilt. That’s the whole pitch of smart beta — and also its trap, because once everyone can choose the same tilt for 15 basis points, the tilt gets crowded.

Welcome to the factor zoo

If a handful of factors are real, how many have been claimed? Far, far too many. In a landmark 2016 review, Harvey, Liu, and Zhu — “…and the Cross-Section of Expected Returns” — catalogued over 300 published factors purporting to explain stock returns, with new ones arriving at a steady clip. The image that stuck is the factor zoo: a sprawling menagerie of “anomalies,” most of which are reskins of a few real animals.

Analogy. Picture a Pokédex with 300-plus entries — but most are palette-swaps of the same dozen creatures. A “new” factor is often value wearing a different hat, or pure data-mined noise dressed up as a discovery.

The multiple-testing problem. Here’s the statistical rot at the core. The conventional bar for “significant” in a single study is a t-statistic above 2.0 (roughly a 5% false-positive rate). But the factor literature isn’t one test — it’s thousands of researchers running thousands of regressions, publishing the winners and quietly burying the losers. When you run enough tests, some will clear $t > 2.0$ by pure luck. With that much data mining, a 2.0 hurdle is hopelessly lax.

Harvey, Liu, and Zhu argue that to account for all the hidden tests, a credible new factor should clear a much higher bar — a t-statistic of roughly 3.0 or more — and even that is a floor, not a guarantee. Their blunt conclusion, echoing the broader replication crisis: “most claimed research findings in financial economics are likely false.”

If you remember one statistic from the zoo, remember the asymmetry: hundreds of factors published, a handful that replicate. The literature is a haystack with a few needles, and the publication process actively manufactures hay.

Out-of-sample decay

Suppose a factor does clear the bar and gets published. What happens to its premium afterward? It shrinks — and it shrinks in two distinct stages. The definitive study is McLean and Pontiff (2016), who took 97 anomalies from the literature and tracked each one’s return after the sample period used in its original paper.

Their two headline numbers, and crucially, what each one means:

1. About 26% lower out-of-sample. Comparing the original in-sample return to the return in the period after the paper’s sample but before publication, the premium fell by roughly a quarter. This drop is mostly statistical: in-sample results are inflated by overfitting and plain luck (you found the factor partly because it looked good on that data), so a fresh, unmined window naturally gives back some of that.

2. About 58% lower post-publication. Comparing in-sample to the period after the paper is published, the premium fell by well over half. The extra decay beyond the 26% — the gap between the two numbers — is the publication effect: once the anomaly is public, arbitrageurs trade it, push prices toward fair value, and arbitrage the premium away.

The logic of the gap is the whole point. Out-of-sample-but-pre-publication decay (26%) is what honest statistics costs you. The additional decay to 58% post-publication is what other traders reading the same paper cost you. McLean and Pontiff confirmed this by showing the post-publication decay was concentrated exactly where you’d expect arbitrage to bite hardest — and that trading volume and correlations in those stocks jumped after publication.

Crowding, capacity, and the cost of trading

Decay isn’t an abstract statistical artifact — it happens through three concrete, brutal mechanisms that any live factor strategy must survive.

Crowding. When a factor becomes popular, capital floods in. Buying cheap (value) stocks en masse bids their prices up, which compresses the very valuation spread that generated the premium in the first place. The more loved a factor, the more its future premium is bid away. Worse, crowded trades are fragile: when everyone holds the same factor portfolio and a shock hits, they all rush for the same exit. The textbook case is the August 2007 “quant quake”: numerous quantitative funds held near-identical factor exposures, and when a few were forced to deleverage, their selling moved factor portfolios against all the others — a synchronized, self-reinforcing unwind that vaporized years of returns in days, despite no fundamental news. Crowding turns a diversifier into a contagion vector.

Capacity. Some factors live in corners of the market that can’t absorb much money. Small-cap and momentum strategies trade in less-liquid stocks and (for momentum) trade frequently; scale them up and your own buying and selling moves prices against you. A strategy that returned 8% on $100 million might return 3% on $10 billion — same idea, but your footprint has eaten the edge. Capacity is why a factor that’s real in a backtest can still be untradeable at size.

Transaction costs. Backtests are typically computed on paper, gross of trading costs. Momentum is the poster child: it requires constantly selling fading winners and buying fresh ones, so its turnover is high, and commissions, bid-ask spreads, and market impact can devour a large chunk of its gross premium. The gross-vs-net gap is where many “great” strategies quietly die.

Worked example — from a significant backtest to a rounding error. Here’s the full lifecycle of a typical factor, with the arithmetic in prose (not in any chart):

Walk through it. The factor showed a 5% gross annual premium with a t-statistic of 2.3 — enough to clear the old $t > 2.0$ bar (so it gets published) but short of the Harvey-Liu-Zhu $t > 3.0$ hurdle (so you should already be suspicious). Apply the typical post-publication decay of about 58%: $5\% \times (1 - 0.58) = 5\% \times 0.42 = 2.1\%$. Then subtract roughly 1.5% per year for turnover and trading costs: $2.1\% - 1.5\% \approx 0.6\%$. A backtest that looked “statistically significant” with a 5% edge becomes, live and net of costs, a rounding error of 0.6% — easily swallowed by fees, taxes, or a single bad year. This is the gap between a paper and a portfolio.

Can you time factors? (Mostly no.)

If factors come and go, surely the move is to time them — pile into value when it’s cheap, rotate to momentum when it’s hot. It’s seductive, and it mostly doesn’t work.

Why timing is so hard. Factors are themselves volatile and mean-revert slowly — value can underperform for a decade (see 2010–2020) before snapping back, far longer than most investors’ patience or career. The most natural timing signal, the valuation spread (how cheap the cheap stocks are relative to the expensive ones), is weak and noisy: it has some predictive power in theory but is swamped by sampling error in practice, and acting on it aggressively often means doubling down right before the pain continues. The weight of the evidence says the same thing: diversify across factors rather than try to time them. A balanced multifactor portfolio harvests several imperfectly-correlated premia and doesn’t need you to call the turning points — which, conveniently, you can’t.

The trade-off, stated honestly. Smart-beta ETFs are genuinely a good thing: they democratize cheap, diversified factor exposure that used to cost active-management fees. But the same accessibility is corrosive — the easier it is for everyone to buy a factor, the more crowded the trade becomes and the more the premium you bought gets compressed. There’s no free lunch even in the discount aisle. The factors most likely to keep paying are the old, theoretically grounded, expensive-to-arbitrage ones — value and momentum have ~century-long, multi-market track records and plausible risk/behavioral stories — not the newest, flashiest entry in the zoo, which is statistically most likely to be the next anomaly to decay to zero. When someone pitches you a brand-new factor, the base rate says: probably noise.

Putting it together

Factor models, taken all the way to the edge, end in humility. Smart-beta ETFs turned factors into cheap, rules-based products — beta you choose, not alpha you’re sold — sitting between passive and active. But the research engine that found a few real factors also spawned a factor zoo of 300-plus published anomalies, most of them statistical mirages born of multiple testing, which is why a credible new factor should clear a much stiffer bar (~$t > 3.0$, an argued adjustment, not a law) than the lax $t > 2.0$. Published anomalies decay in two stages — about 26% lower out-of-sample (overfitting unwinding) and about 58% lower post-publication (arbitrage on top) — and the live premium gets further chewed by crowding (the 2007 quant quake), capacity limits, and transaction costs, so a 5%, t = 2.3 backtest can net 0.6%. Timing factors mostly fails because they mean-revert slowly and timing signals are noisy, so the verdict is diversify, don’t time. The durable factors are the old, grounded, hard-to-arbitrage ones; the newest zoo entry is, until proven otherwise, probably noise.

That’s the full arc of factor models — from CAPM’s single broken beta to a disciplined, skeptical toolkit. You can now read a fund’s factor fingerprint, separate genuine alpha from chosen beta, run and interpret a factor regression, and treat the next shiny anomaly with exactly the suspicion it has earned. The deepest lesson of the whole course is the one this lesson hammered home: in markets, the moment an edge is known, it starts to disappear — so the durable edges are the few that are real, grounded, and hard to copy, and the right posture toward everything else is principled doubt.