Strategy decay: every edge dies, and how to notice early

Why edges die

An edge is an inefficiency. Markets exist to remove inefficiencies. The mechanisms differ. The direction never does.

Crowding. A pattern works, gets noticed, gets copied. Every new copier moves price toward efficient before you collect. The early ones profited. The late ones pay costs to stand in line. Your edge did not break. It got shared until the shares were worthless.

Regime change. The market state the strategy was built for ended. The trend strategy meets the chop. Nothing about the strategy changed, and that is precisely the problem. See market regimes.

Structural change. A new venue, a new contract, a new fee schedule. The mechanism the strategy harvested no longer exists in that form. No parameter refresh resurrects a mechanism the market removed.

Sample luck running out. The backtest window happened to be friendly. Live trading reverts to the long-run average, which is lower. The strategy was never as good as its history. The history was just generous.

The fit revealing itself. The harshest one. The edge never existed. The backtest was a curve fit, and live trading is the truth arriving on schedule. See overfitting.

What decay looks like from the inside

Decay almost never arrives as one terrible day. It arrives as a pattern, and every piece of the pattern is individually dismissible.

• Sharpe drifts down. 1.2 in the backtest. 0.9 the first month. 0.5 by the third. Each step looks like noise. That is how decay stays cheap to ignore and expensive to have ignored.
• Drawdowns last longer. The wins still come. The recoveries stretch from days to weeks.
• The win rate erodes. Same trades, same sizes, 45 percent instead of 55.
• The signals thin out. The market offers fewer of the setups the edge needs.
• Trade quality slips. Wins shrink against costs. The per-trade margin compresses while the headline PnL still looks survivable. This one often moves first. See trade quality.

Decay or noise: the central question

Every strategy has bad weeks. Killing systems on noise costs as much as holding them through decay. So the discrimination matters, and feelings are terrible at it.

One bad week is noise. Three bad months with the same metric pattern is decay. Losses in a regime the strategy was never built for are not decay; they are the deal you accepted, and the regime decomposition will say so. Losses in the regime the strategy is supposed to love are decay. A live-to-backtest gap inside 20 percent is noise and model error. Beyond 40 percent, decay or the fit revealing itself.

The principled answer is statistical: test the rolling live performance against the backtest’s distribution and ask whether this divergence happens by chance. Eyes cannot do this. Eyes see decay that is not there at the bottom of every normal drawdown, and miss decay that is there for as long as hope holds out.

The four responses

In the order to consider them:

• Recalibrate. Re-optimize on recent data. If the logic is sound and the market shifted, a parameter refresh can restore the edge.
• Cut size. Uncertain? Halve it. You keep collecting evidence, you lose less while deciding, and no decision made smaller ever ruined anyone.
• Pause and watch. Some strategies revert when their regime returns. A month on the bench is often the cheapest information available.
• Retire it. Recalibration failed and the regime did not change. The edge is gone. Free the capital. This should be a planned outcome with a written trigger, not a crisis with a midnight decision.

Crypto runs the clock faster

Retail crowds copy patterns quickly. Venues mutate quickly: new perps, new chains, new fees. Funding regimes turn in months. And the institutional wave keeps arriving, arbitraging away the easy edges that survived only because nobody serious was looking. An edge’s life expectancy here is shorter than in equities. Plan accordingly.

The pipeline is the strategy

One strategy is fragile. A pipeline is durable. Three habits separate the operators from the gamblers. Write a sunset condition with numbers before deploying, on a calm day, because the bottom of a drawdown is the one place those numbers cannot be chosen rationally. Keep a validated replacement warm; the day you notice decay should be a deployment day, not a brainstorming day. And treat retirement as success. A strategy that paid for eighteen months and was sunset on schedule did its job. See the psychology of running a system.

How Edgecraft handles this

Edgecraft offers tracking live performance statistically against the backtest envelope, continuously, so the decay-or-noise question gets answered with a test instead of a feeling. Drifting past the threshold raises a flag before losses compound. The recalibration workflow re-optimizes on fresh data and reports whether the edge is restorable or finished. And the strategy library keeps candidates ready, so retirement is a transaction instead of an emergency.