Why edges fade: how to tell a bad patch from a strategy that stopped working

Why do trading edges fade?

An edge is a repeatable reason a set of trading rules makes money on average, the idea Vol 1 starts from.

It was never real. McLean and Pontiff (2016) followed 97 variables that academic papers had shown to predict stock returns. Portfolio returns were 26% lower in the years after each study's sample ended but before publication, and 58% lower after publication. They call the first drop "an upper bound estimate of data mining effects." Part of what looks like decay is the overfitting in Vol 2, which testing on unseen data in Vol 3 is built to catch.

Other traders found it. McLean and Pontiff estimate "a 32% (58% - 26%) lower return from publication-informed trading." The falls were larger for predictors with higher in-sample returns or cheaper to arbitrage. In one documented episode, Khandani and Lo's working paper finds that an unwind, a rapid closing of positions, of similarly built long/short stock portfolios "began in July 2007 and continued until the end of 2007."

Costs ate it. Chen and Velikov (2023) studied 204 stock market anomalies, published return patterns. Net of effective bid-ask spreads, post-publication effects and the trading technology of the 2000s, "the average anomaly's expected return is a measly 4 bps per month," 0.04%.

The market changed state. Ang and Timmermann describe financial markets that "often change their behavior abruptly", where "the new behavior of financial variables often persists for several periods after such a change." Daniel and Moskowitz find that momentum strategies "can experience infrequent and persistent strings of negative returns" in "panic" states after market falls.

Not all of it fades. Jensen, Kelly and Pedersen (2023) report that most published asset pricing factors can be replicated and work out of sample, on data covering 93 countries, and call "a positive but attenuated post-publication alpha", a smaller edge that survives, the expected outcome. Outside the US, Jacobs and Muller (2020), as quoted there, found no reliable post-publication decline.

Why is decay so hard to see?

In Figure 1, two simulated accounts run on the same daily shocks at 10% yearly volatility, with no costs. One keeps a true Sharpe ratio of 1.0. In the other, the Sharpe ratio falls from 1.0 at year 4 to 0 at year 10. Their average daily returns never differ by more than 0.04%, against a typical daily move of 0.63%, and both had their worst day on the same day: -3.93% and -3.95%. The fading account fell 10% behind only in year 7.6. McLean and Pontiff call post-publication decay "a slow process" and add: "We are unaware of theories on how long the decay should take or on the functional form of the decay."

Is it a bad patch or the end of the edge?

A Sharpe ratio measured on a short record is noisy. For independent normal returns, Lo (2002) gives its standard error, the typical gap between a reading and the true value; applied to 12 months of daily returns at a Sharpe ratio near 1, it comes to about 1.0 in yearly units, as large as the whole distance from 1.0 to 0. Rules with a true Sharpe ratio of 1.0 therefore land anywhere from about -0.65 to 2.65 over a year, 9 times in 10, and lose money over 12 months about 16% of the time.

Across 2,000 simulated 12-year paths with a true Sharpe ratio of 1.0, the 12-month readings came close to the formula; the range they fill 9 times in 10 is the band in Figure 2. A falling rolling Sharpe ratio is the common way to watch for decay ("If Sharpe is getting worse over time, the strategy is decaying," in the Trading Strategy glossary), yet at 12 months much of that movement is noise.

In the first panel, with the edge unchanged, the reading dropped below the band in year 7, was back above its middle 7.6 months later, fell again to -2.06 in year 8.7. In the second panel, the fading account from Figure 1, the reading sat below the band's middle in 90% of months in years 9 to 12. Around year 9 both readings are low; later averages differ, though individual readings still overlap.

Robot Wealth, a trading blog, describes a Sharpe-1 bond strategy's rough 2023-2024 stretch and estimates that, after the author stopped trading it, the strategy would have returned about 15% at a Sharpe of 1.8 after costs in 2025.

How long can a normal drawdown last?

Time under water is how long an account stays below its previous high. Bailey and Lopez de Prado (2015) study a different quantity: when the 5th-percentile cumulative profit, the running total that 1 path in 20 falls below, returns to zero under independent normal cashflows. Their formula, (1.645 / Sharpe ratio) squared, gives 2.7 years at Sharpe 1.0 and 10.8 at 0.5. It does not give the 95th percentile of drawdown lengths, the length that 1 drawdown in 20 exceeds. Along that 5th-percentile curve, recovery takes three times as long as the fall to the lowest point, their "triple penance rule."

The simulation measures the longest stretch under water in a 10-year record. One record in 20 had a stretch of at least 8.25 years at a Sharpe ratio of 0.5 and 4.17 years at 1.0. At a Sharpe ratio of 0.5, 22.9% of records spent 5 years or more under water, with the edge intact the whole time.

Rej, Seager and Bouchaud (2017) find "a 5% chance for the drawdown of a ten-year process with SR = 0.5 to last 7 years or more!" and turn it into a test: "a drawdown that lasts too long or one that is too deep should lead to a downward revision of the assumed Sharpe ratio of the strategy." They also note that the test "does not allow to distinguish" a Sharpe ratio that was estimated wrongly in the first place from bad luck.

How long until the numbers can tell?

Two standard errors is the usual 95% yardstick from Vol 1, and a Sharpe ratio measured on n years of new data has a standard error of about 1 divided by the square root of n. For an edge with a Sharpe ratio of 1.0:

1. If the edge vanishes, the true value moves from 1.0 to 0, a distance of 1. Two standard errors shrink to 1 at n = 4 years.
2. If the edge halves, the distance is 0.5. Two standard errors shrink to 0.5 at n = 16 years.

A measured Sharpe ratio wobbles around its true value, so the simulation also timed the wait directly on 2,000 paths: a known change after 5 years, and a flag raised whenever the Sharpe ratio of the new data alone sat two standard errors below 1.0.

At year 4, 48% of vanished-edge paths were flagged. The median wait for a flag that persisted through year 30 was 5.4 years. For a halved edge it was 16.3 years among paths flagged at year 30; 23% were unflagged then. Over a single month with zero edge, Robot Wealth's simulation gets the same answer: "The t-test finds nothing."

Can a market regime explain a losing stretch?

A regime is a stretch where the market behaves in a similar way, such as trending or choppy. Rules built for one state can lose for months while the market sits in another.

In Figure 4, a simple trend rule, long after a 20-day rise and short after a 20-day fall, made 60.9% in the first trend, lost 21.9% through the choppy stretch and made 49.0% when the trend came back, with the same rules throughout.

The strip under the price is a label from Perry Kaufman's efficiency ratio over 40 days: the net price move divided by the sum of every bar-to-bar move, near 1 for a straight line and near 0 for noise. Built from past prices, it switched to choppy 38 trading days after the chop began and back 40 days after the second trend began.

A regime label is a hypothesis about a losing stretch, and a plausible one can mislead. The Robot Wealth post describes its author's own mistake: "because I could construct a reasonable narrative about why the edge might have disappeared, I convinced myself that my judgment was adding value. Turns out I was just pattern-matching on noise."

What does the research say about edge decay?

• Overfitting predicts which strategies decay. In Falck, Rej and Thesmar's preprint on 72 published strategies, the Sharpe ratio "declines by about one half after the results are published." The year of publication alone explains 30% of the variation in decay, and measures of overfitting add another 15%. Arbitrage-related measures add little.
• Costs take most of what remains. After spreads and post-publication effects, Chen and Velikov (2023) put the average anomaly at about 4 basis points a month, even though Jensen, Kelly and Pedersen (2023) find most factors replicate.
• Normal drawdowns are longer than people expect. Rej, Seager and Bouchaud: "both managers and investors tend to underestimate the length and depth of drawdowns consistent with the Sharpe ratio of the underlying strategy." Bailey and Lopez de Prado note that "investment firms typically set less strict stop-out rules to portfolio managers with higher Sharpe ratios."

Common mistakes

• Reading one bad year of a Sharpe-1 strategy as proof. Such rules lose money over 12 months about 16% of the time with nothing broken.
• Treating the backtest's worst drawdown as the worst normal one. In the simulation at a Sharpe ratio of 1.0, the median worst drawdown was -11.1% in 3-year records and -18.4% in 20-year records, and the 1-in-20 worst went from -20.0% to -28.3%.
• Trusting a story about why the edge died. A reasonable narrative can be built for noise too.
• Re-tuning the rules after a bad patch. One more round of the overfitting in Vol 2.
• Taking a rule like "20+ trades" at face value. Google's AI Overview for this question offers "5% or more below baseline over 20+ trades" with no source. At a 40% win rate, a win rate measured over 20 trades has a standard error of about 11 points.