1.10.5 Failure Modes & Real-World Practice

Basic Prerequisites: 03 · Crowding & Capacity and 04 · Post-Publication Decay.

---

1. Intuition & Practical Objective

The previous pages derived how a factor premium can be estimated, sized, and eroded. This page assembles the operating manual: the six ways factor investing actually loses money, each named precisely, tied to a first principle, and - where possible - given a number. The objective is not cynicism. It is knowing exactly where the elegant arithmetic stops describing the world.

The six failures, in one line each:

1. The diversification illusion - many “factors” that are highly correlated are one factor wearing several names; the Sharpe you thought you diversified away was never there.
2. Crowding & the liquidity black hole - a forced unwind of a crowded factor has no counterparty; the tail is far fatter than the risk model says.
3. Post-publication decay - a published premium loses ~35% of its in-sample alpha, and the decay is fastest where arbitrage is cheapest.
4. Capacity & trading costs - net alpha is concave in AUM and dies at a break-even that scales with the square of the gross premium.
5. The factor zoo / multiple testing - with hundreds of candidates, hundreds of chance discoveries; the right hurdle is $t>3$, FDR, or a deflated Sharpe.
6. Factor timing is a high-variance, low-IC bet - timing has breadth ≈12, so it needs an IC ≈0.1 to matter, and in-sample-optimal timing weights routinely collapse out-of-sample (developed on 06 · Advanced Extensions).

The one-sentence essence. “Factor investing fails in six predictable ways - correlated ‘factors’ that are one bet, crowded unwinds with no counterparty, publication decay of ~a third, capacity that dies quadratically in the premium, multiple testing that manufactures discoveries, and timing that needs an IC three times stronger than stock selection - and a practitioner’s skill is measuring which one is active right now.”

---

2. Mathematical Ground Truth & Derivations

Failure 1 - the diversification illusion, quantified. Take $N$ factor strategies, each with expected alpha $\mu$ and residual volatility $\sigma$. Their residuals are not independent: suppose the average pairwise residual correlation is $\rho$. Then an equal-weighted portfolio of the $N$ strategies has alpha $\mu$ and volatility

$${\sigma }_{\text{port}}=\sigma \sqrt{\rho +\frac{1-\rho }{N}},$$

so its appraisal ratio (information ratio) is

$$I{R}_{\text{port}}=I{R}_1\sqrt{\frac{N}{1+(N-1)\rho }},I{R}_1=\frac{\mu }{\sigma }.$$

Two limits matter. As $\rho \to 0$, $I{R}_{\text{port}}\to I{R}_1\sqrt{N}$ - the diversification we hope for. As $\rho \to 1$, $I{R}_{\text{port}}\to I{R}_1$ regardless of $N$ - twenty factors become one. This is why Cochrane insisted that covariance, not the mean, is the central object: “if the value firms decline, they all decline together.”

Failure 2 - crowding, the flow-shock term the risk model omits. A standard risk model writes $\Sigma =B\Omega B^{\prime }+D$ with $D$ diagonal, i.e. it treats idiosyncratic and flow-driven moves as independent noise. In a crowded unwind they are perfectly aligned: everyone sells the same names. The correct structure has an extra term for the common flow shock,

$${\Sigma }_{\text{crowded}}=B\Omega B^{\prime }+D+\zeta s{s}^{\top },s=\text{the crowded bet’s direction},$$

and $\zeta$ is time-varying - it is small in normal times and enormous in a deleveraging. No historical covariance matrix captures that, so the mitigation is a stress overlay, not a bigger $\sigma$: scenario impact, participation caps, liquidity-adjusted VaR.

Failure 3 - decay, restated as a half-life. If a premium decays exponentially at rate $\lambda$, $g_t=g_0{e}^{-\lambda t}$, the half-life is $t_{1/2}=\ln 2/\lambda$. McLean–Pontiff’s ~35% post-publication decline over their post-publication window implies a long but finite half-life; the practitioner’s rule is to apply a haircut of roughly $1-{\hat{d}}_{\text{post}}$ to every backtested alpha before evaluating it. With a 35% haircut, an in-sample 5%/yr is worth 3.25%/yr - and if the factor is liquid and low-idio-risk, haircut more (their cheap-to-arbitrage group decayed ~49%).

Failure 4 - capacity. From 03: $g_{\text{net}}(A)=g-\tau 2\lambda \sigma \sqrt{A\tau /(252ADV)}$, with break-even $A^{\star }\propto ADV{g}^2/(\lambda \sigma {)}^2$. Halving the gross premium quarters the capacity.

Failure 5 - the multiple-testing floor. Under the null with $M$ tests, the expected false-discovery count is $0.0455M$; a $t>3$ hurdle is the standard remedy; BH-FDR controls the fraction rather than the count.

Failure 6 - timing breadth. $I{R}_{\text{timing}}=IC\sqrt{12}$; to reach IR 0.5 you need a monthly timing IC of ~0.14, on a single highly volatile bet. (Developed on page 06.)

---

3. Computational Implementation - the two failures you can see in a backtest

Stdlib only. Experiment A shows eighteen of your twenty factors were one factor; Experiment B is the crowding monitor every factor book should run - the moment pairwise correlation creeps up, the risk model’s diversification forecast is already wrong.

Twenty uncorrelated factors give an IR of 1.50 (closed form 1.55) - genuine diversification, a near-5× improvement over one factor. Twenty factors with 85% residual correlation give an IR of 0.29 (closed form 0.37), which is indistinguishable from holding a single factor. Eighteen of them bought you nothing but fees. A real factor book routinely holds “value”, “quality”, “low-vol”, and “profitability” simultaneously; if their pairwise correlation is 0.85 they are one bet, and the portfolio’s advertised diversification is fiction.

Both factors keep the same mean and the same volatility throughout - nothing in either factor’s own statistics changes. What changes is their correlation, from +0.07 to +0.83. The consequences: the book’s realized volatility is 1.31× the risk model’s forecast; the Sharpe collapses from 0.88 to 0.09 (the premium is unchanged, but it is no longer diversified); and the worst month deepens to −8.18%. This is the mechanism that turns a market-neutral book into a one-way bet without anyone changing a position - and it is only visible if you monitor pairwise factor correlation, not factor volatility.

---

4. Failure Modes & First-Principles Breakdowns (numbered)

1. The diversification illusion. As Experiment A shows, $I{R}_{\text{port}}=I{R}_1\sqrt{N/(1+(N-1)\rho )}$ is bounded by $I{R}_1$ as $\rho \to 1$. Practice: estimate the residual-correlation matrix of your factor sleeves, orthogonalize (Gram–Schmidt or PCA) before combining, and count eigenvalues, not sleeves. If the correlation matrix’s second eigenvalue is tiny, you own one factor.
2. Crowding & liquidity black holes. August 2007 (Quant Quake) and March 2020 (the value crash) are the canonical episodes: forced deleveraging cascades into same-side liquidations across the whole factor, producing unprecedented drawdowns in supposedly market-neutral books. The factor’s modeled beta-risk says “low”; the realized tail says otherwise - because everyone holds the same long-short. Practice: monitor cross-factor correlation, valuation spreads, short interest, and participation; hold a stress scenario, not just a covariance matrix.
3. Post-publication decay. ~35% average haircut, larger for cheap-to-arbitrage characteristics (~49% in the low-idio-risk group), smaller for costly ones (~17%). Practice: haircut backtested alphas before sizing; re-estimate the premium on post-publication data only; prefer factors that are hard to trade - they decay slower.
4. Capacity & trading costs. Net alpha is concave in AUM, and $A^{\star }\propto g^2$. A factor with 2%/yr gross alpha has roughly a quarter the capacity of one with 4% (≈$158bn vs ≈$630bn for the §03 calibration). Practice: build the impact model into the backtest, cap participation at a few percent of ADV, and report capacity alongside Sharpe.
5. The factor zoo / multiple testing. Expected false discoveries $=0.0455M$; use $t>3$ or BH-FDR; state the number of configurations you tried. See Backtesting Hygiene & Deflated Sharpe.
6. Factor timing is a high-variance, low-IC bet. Timing breadth ≈12, so an IR of 0.5 needs a monthly timing IC ≈0.14. In-sample-optimal timing weights collapse out-of-sample (page 06 quantifies the collapse). Practice: treat timing as a small overlay (a tilt of ±0.25–0.5× the factor’s normal risk budget), validate with purged cross-validation, and never size it as if it were a stock-selection signal.
7. The meta-failure: every one of these is a time-varying quantity. Correlations, decay rates, capacities, and premia are not constants to be estimated once; they are state variables. A factor programme is a monitoring system, and the practitioner’s edge is in noticing the regime change before the P&L does.

---

5. References

• Khandani, Amir & Lo, Andrew, “What Happened to the Quants in August 2007?” (JIM, 2007)
• McLean & Pontiff (JF, 2016)
• Almgren et al. (Risk, 2005)
• Ilmanen, Expected Returns (2011)
• Cochrane, (JF, 2011) §II
• Grinold, Richard & Kahn, Ronald, Active Portfolio Management
• Taleb, Nassim Nicholas, Dynamic Hedging
• Sibling practitioner page: the construction folder’s failure page - multicollinearity in $X^{\top }X$, the factor zoo, and the same crowding episode seen from the model side.

---

6. Connected Graph Bridges

• Back: 04 · Post-Publication Decay · 03 · Crowding & Capacity · Index Hub
• Forward: 06 · Advanced Extensions: Factor Timing
• Risk & construction: Quantitative Risk · Portfolio Optimization (orthogonalized factor sleeves, capacity constraints)
• Discipline: Backtesting Hygiene & Deflated Sharpe · Regime Detection (detecting the state change)
• The factor family: Momentum (momentum crashes) · Fundamental Multi-Factor Models