Factor Crowding and Decay: Why Every Factor Dies and How to Extend Its Life

The premium of any published factor will eventually be arbitraged to zero. The only question is whether you’re early enough in the cycle.

This is the dirty secret of factor investing. The academic literature describes factors with Sharpe ratios of 0.5–1.0 that persist for decades. The reality: most of those premia shrank by 30–60% within ten years of publication, because quant funds read the same papers you did and deployed capital on the same trades.

McLean and Pontiff (2016), Does Academic Research Destroy Stock Return Predictability?, Journal of Finance, put numbers on it. They find anomaly-portfolio returns are roughly 26% lower out-of-sample — using only data after each anomaly’s own original sample period — and roughly 58% lower once the paper is published and widely cited. The additional ~32 percentage points between those two figures is the money that arrived after the paper went public. ISSN 0022-1082.

The gap between those two numbers carries the meaning. Even if nobody had read your paper, out-of-sample returns would still fall; that portion is data-mining bias giving itself back. The extra decay after publication is crowding. Crowding is the part worth fighting, because it’s the only part connected to anyone’s behavior.

Four Phases of Decay

A factor’s life splits into roughly four stretches. The timescales are order-of-magnitude estimates, not a schedule.

Discovery (years 0–3). The academic paper appears. The premium is large, often Sharpe above 1.0. Almost nobody is trading it. The professor who found it may have a personal account running the strategy.

Institutional adoption (years 3–10). Quant funds read the paper and each builds a variant. Capital on the factor grows from near zero to the tens-of-billions range. The premium starts shrinking and Sharpe drops to 0.5–0.8. The early movers still make money; the late movers crowd into a smaller edge.

Maturity and crowding (years 10–20). The factor is well known and capital in factor products reaches the hundreds of billions. The marginal trader earns below the cost of capital. The strategy still works for anyone with a structural advantage — lower transaction costs, faster execution, better financing. For everyone else it’s noise.

Noise (year 20 onward). Unless the factor has structural reasons to persist, like transaction costs that keep retail out or deeply rooted behavioral biases, the premium is mostly gone. The strategy survives as a marketing claim more than a profitable trade.

The value factor (HML) has walked roughly this path from 1992 to today.

Three Mechanisms

Decay has no single cause. Three mechanisms chase each other.

Arbitrage is the standard efficient-markets story. Once a factor is published, the prices of assets that would benefit from the trade rise and the prices of assets that would suffer fall, and the mispricing closes. In a frictionless market all known factors would be arbitraged to zero instantly. We don’t live in one, so arbitrage takes years — but it does happen.

Crowding is arbitrage’s side effect, and it bites harder. Too many traders on the same side of the same trade. It works fine when it’s 50 funds with $10B each; it breaks when the marginal position is held by someone who’ll sell at the first sign of trouble. What crowding produces isn’t low returns, it’s fragility: when the trade goes against you, liquidity disappears, spreads widen, correlations spike, and you can’t get out at the price you expected. Momentum in 2009 was the canonical crowded unwind; value in 2020 was another.

Behavioral change is the slowest and the broadest. Institutional investors learn about factors, and pension funds, endowments, and family offices start tilting toward quality, low-vol, or value. That changes demand for those assets, which changes prices, which shrinks the premium. Institutional surveys over the past decade point the same way: a large share of big institutions now carry explicit factor tilts in their policy portfolios. That’s trillions of dollars chasing the same four to six factors.

Note that these three relate to you differently. Arbitrage you can do nothing about. Behavioral change you simply absorb. Crowding is the only one you can observe in advance and size against — which is what the next section is for.

Crowding Isn’t Directly Observable, Only Its Proxies

You can’t see how much AUM sits on a given strategy. Everything you can see is a proxy, and the most reliable proxy is your own factor’s predictive power.

If a factor had an IC (information coefficient) of 0.10 historically and now reads 0.03, the signal is weaker — and whether the cause is crowding, regime change, or a degraded implementation, the conclusion is the same: your edge is smaller than it was. In Arithmion you compute this through evaluate_factor:

from apexquant.notebook import evaluate_factor

report = evaluate_factor(bars, "momentum_60", quantiles=5, periods=(21,))
ic = report["ic_summary"]
ic["rank_ic_mean"], ic["rank_icir"], ic["rank_ic_t_stat"]

PROMPT

Use Arithmion’s notebook.evaluate_factor to get a momentum factor’s IC summary: rank IC mean, ICIR, and t-statistic.

To see decay rather than level, rerun the report over rolling windows and compare recent windows against the long-run mean. One statistical trap is worth knowing here: when the forward returns in periods span multiple bars, the IC series is autocorrelated, and a naive ICIR × √n inflates significance. The report’s rank_ic_t_stat already applies a Newey-West correction with bandwidth set to the primary period minus one, so judging “has the IC really fallen” should lean on that rather than on dividing the mean by the standard deviation yourself.

The second proxy is correlation between factors. When many factors become the same trade, inter-factor correlations rise. Value and quality used to be uncorrelated; in the 2010s they became strongly correlated, since both amount to “long profitable cheap companies”. If your “diversified” four-factor portfolio is really 80% one trade, the diversification is fake. This step needs no special API — once you have factor returns, it’s a pandas operation:

rolling_corr = factor_returns.rolling(60).corr()
latest = rolling_corr.loc[factor_returns.index[-1]]
worst_pair = latest.where(~np.eye(len(latest), dtype=bool)).abs().max().max()

PROMPT

Use pandas to compute a 60-period rolling correlation matrix across several factor return series and pull the largest off-diagonal correlation from the latest period.

The third proxy lives outside your data: fund flows. You can’t see AUM in “the value factor”, but you can see AUM in value ETFs (IVE, VBR, IUSV and their peers), factor mutual funds, and quant equity flows. That data comes from market and fund-data vendors rather than the platform — which is also its weakness as a signal: it lags, its definitions are inconsistent, and ETFs only cover part of the trade. Treat it as directional corroboration, not a trigger. The rough rule: if capital on a factor doubles inside two years, expect premium compression.

The fourth signal is the bluntest and the latest. Momentum fell roughly 50% over twelve months in 2009; value drew down to the −40% range within a year in 2020; low-vol had a milder version in Q4 2018. A sudden, severe factor drawdown is usually a crowded unwind. Two responses are available at that point: avoid the factor for the following 6–12 months, or size it down.

The Value Factor: One Complete Sample

Value (HML, high minus low book-to-market) is the most complete decay case available, because its whole curve has run.

Fama and French found it in 1992. The documented premium was roughly 5% per year from 1963 to 2007, and the trade fits in one sentence: long cheap stocks, short expensive ones, Sharpe around 0.6–0.8.

What followed, in order. Through the 1990s, AQR, DFA, LSV, and Vanguard all built value strategies and capital in value equity grew from near zero to the tens of billions by 2010. In 2007–2009 the financial crisis handed value one of its worst drawdowns in modern history, as bank stocks — the ultimate “value” trade — collapsed. From 2010 to 2019 value underperformed growth for almost a decade and the annualized premium shrank to the 1–2% range, with a brief 2017 comeback that died again. From 2020 to 2022, as rates rose from zero, value finally outperformed substantially. Then in 2023–2024 the premium shrank again as growth stocks recovered on the AI narrative.

For someone who started running value in 1995, those decades were genuinely profitable, landing near the magnitude the literature documented. For someone who started in 2010, it was probably barely positive after costs. Same factor, same rules; the only difference is when you arrived.

Being five years late can cost you most of the premium. That’s the most practical conclusion in this article, and it explains why everything in the next section is really an answer to one question: I’m late, now what?

What’s Left If You’re Late

Four routes, with very different payoffs and barriers.

Trade a better variant. Book-to-market is one measure of value, and a fairly old one. Enterprise multiple (EBITDA / EV), free cash flow yield, and earnings yield are all candidates, and the “improved value” composite built from them has outperformed classic HML in the literature. The same logic applies to other factors: the first published version is rarely the best version, and by the time everyone trades it the academic literature has long since moved to better specifications. This is the cheapest route, and the most realistic one for retail and small institutions.

Diversify across factors. A four-factor blend has lower drawdowns and steadier returns than any single factor: when value underperforms, momentum or quality may hold up, and when momentum crashes, value often rallies. Be clear about what it doesn’t do, though — diversification reduces how painful the path is, it doesn’t eliminate decay. Every individual factor in the blend keeps decaying on its own schedule; you’ve stacked several declining curves, not repaired any of them.

Be faster. If you can rebalance daily while competitors rebalance monthly, you capture more of the premium; if your costs are 5 bps against their 20, your net Sharpe is higher. This is a genuine structural advantage, and it has to be bought with money and infrastructure. Most retail and small institutional players are not competitive here, and recognizing that beats spending into it.

Go where it’s less crowded. US large cap is the most crowded market in the world; every factor strategy is running on SPX names. Emerging market equities, small caps, less-liquid credit, and certain commodity futures have far fewer participants, so the factor premium is larger and more durable. The costs are on the label: lower capacity, harder data, higher trading costs. Which makes this a route for allocating part of a book, not for relocating the whole thing.

Worth noticing: none of the four is “find a new factor”. That isn’t an accident, and the last section explains why.

Turning Decay Into a Continuous Measurement

Running factor strategies calls for a monitor that stays on, not a one-time due-diligence pass. The list of things to watch is short:

1. Rolling IC of each factor over a 36-month window
2. Rolling correlations among your own factors
3. Fund-flow proxies in factor ETFs and quant funds
4. Drawdown alerts when any factor drops more than 15% in six months
5. Regime indicators (VIX, credit spreads, rate of change)

The first two you can compute entirely inside the platform (evaluate_factor plus pandas is enough). The third needs external data. The last two are your own risk parameters. The platform ships no general “factor monitor” component, deliberately: the shape of that thing depends heavily on which factors you run, at what frequency, and what size of drawdown you care about, and a generic component would mostly force you to accommodate its assumptions.

The point is that it isn’t a set-and-forget system. Factor decay happens continuously; no event fires to tell you it started. The factors that worked last year may not work next year, and the ones that don’t work now may come back — value from 2020 to 2022 is the living example. So the output of a monitor shouldn’t be alert-or-no-alert. It should be a curve you glance at every quarter.

The Honest Take

If you’re starting quant research today, the edge isn’t in finding new factors. Finding new factors has been done at least two hundred times over, and the newly published ones mostly won’t survive a multiple-testing correction. The edge is elsewhere: combining known factors more intelligently, executing more cheaply and faster, trading where fewer people are, and noticing your own factor’s decay earlier than the next person.

The heyday of factor investing — discovering HML in 1992 and trading it for 30 years — is over. The new game is incremental edge in a crowded market: harder, less alpha, more infrastructure.

I don’t find that depressing. It just means the work is harder and the rewards are smaller, and both of those are good news for anyone doing it seriously, because what it filters out is the people who assumed the premium was permanent.

Concretely, the order runs roughly like this. Use the platform’s factor evaluation to establish a baseline for four to six low-correlation factors, so you know what each one’s IC and ICIR read today. Attach the five monitors above and let them refresh quarterly. Carve out a modest slice of the book — on the order of 10–20% — for smart variants of established factors, and a comparable slice for less-crowded markets. Adjust the allocation quarterly on crowding signals rather than on mood. And manage expectations last: net of costs, factor Sharpe ratios in the 0.4–0.7 range are normal, not the 0.8–1.2 the academic papers suggest.

The premium is real. It’s smaller than advertised, and it will keep decaying. The people who manage that decay well stay in business a long time.

References

• McLean, R. & Pontiff, J. (2016). Does Academic Research Destroy Stock Return Predictability?. Journal of Finance. ISSN 0022-1082.
• Chordia, T., Subrahmanyam, A., & Tong, Q. (2014). Have Capital Market Anomalies Attenuated in the Recent Era of High Liquidity and Trading Activity?. Journal of Accounting and Economics. ISSN 0165-4101.
• Fama, E. & French, K. (1993). Common Risk Factors in the Returns on Stocks and Bonds. Journal of Financial Economics. ISSN 0304-405X.
• AQR Capital Management. Can You Beat the Market? Factor Investing. Research paper, 2019.
• Harvey, C., Liu, Y., & Zhu, H. (2016). …and the Cross-Section of Expected Returns. Review of Financial Studies. ISSN 0893-9454.