Every trading strategy that works will eventually stop working. This is not a pessimistic observation — it is the correct prior for systematic trading. Most edges are regime-dependent, exploitable, or fragile to structural changes in the market. The traders who last are not the ones who find a permanent edge. They are the ones who recognize decay early and rotate out before it costs them.

The mistake is treating an edge as a fixed asset rather than a depreciating one. When a strategy begins underperforming, most traders' first response is to wait. The backtest showed it worked. The logic is sound. This must be a drawdown. That response is correct some percentage of the time and catastrophically wrong the rest. The framework that determines which case you are in is what this article is about.

The Three Causes of Decay

Not all edge decay is the same. Three distinct mechanisms cause strategies to stop working, and the diagnosis determines the correct response.

Market regime change is the most common cause. Most edges are not universal — they are regime-dependent. A mean reversion strategy captures edge when prices oscillate around a stable mean. In a trending regime, the same strategy repeatedly fades moves that continue. The strategy did not break. The market it was designed for is no longer present.

Crowding is the second cause. As a strategy becomes known — through published research, shared code, or widespread adoption — more participants exploit the same edge. Competition compresses the spread. Slippage increases because the same signals are triggering the same trades at the same time. The edge arbitrages itself away. This happens faster now than it did twenty years ago. Execution speed, low-cost data feeds, and accessible backtesting tools mean that a documented edge can be crowded within months.

Structural market changes are less frequent but permanent. Regulation NMS in 2007 reorganized how US equities trade. HFT proliferation changed the depth and behavior of limit order books. The growth of options markets created hedging flows that did not exist before. Strategies built on microstructure patterns from 2003 were operating on a market that no longer existed by 2010. When structure changes, edges that depended on the old structure die. They do not recover.

Regime-Dependent Edges

Understanding regime dependency is the first line of defense against decay. Mean reversion works in range-bound markets where prices revert to equilibrium after shocks. It fails in trending markets where momentum carries prices through equilibrium and beyond. Momentum strategies have the inverse problem — they perform in trending regimes and bleed in choppy ones.

A simple regime filter uses ADX (Average Directional Index). ADX above 25 signals a trending regime; below 20 signals a range-bound one. A mean reversion strategy running during ADX > 25 periods should be flagged for review — not necessarily turned off, but watched. Rolling realized volatility serves a similar purpose: when volatility compresses into a narrow band, mean reversion tends to be more reliable. When volatility expands sharply, trending dynamics often take over.

Neither of these filters guarantees anything. They give you the vocabulary to describe what the market is doing, which is necessary before you can evaluate whether your strategy's environment has changed.

Crowding and Arbitrage

CTA trend-following is the most cited example of crowding. The strategy — buy what is going up, sell what is going down across a universe of liquid futures — is well documented, well understood, and practiced by hundreds of systematic funds. Over the past decade, capacity in the strategy has expanded and measured performance has degraded from historical benchmarks. The edge has not disappeared. There is still trend-following alpha. The spread has been arbitraged to the point where it requires more sophisticated execution and larger diversification to extract.

At the retail level, crowding is more severe. A mean reversion strategy that worked on 5-minute bars in 2018 may have been adopted by enough algorithmic traders by 2021 that the statistical properties of the signal changed. The autocorrelation structure that the strategy relied on was competed away. From a backtest against pre-2020 data, the strategy looks fine. From live trading in 2022, the fills are worse, the signals are noisier, and the Sharpe has collapsed.

The crowding mechanism is self-obscuring. The strategy does not announce that it is crowded. Performance degrades gradually, which looks like a prolonged drawdown, which feels like something to wait out.

Structural Market Changes

Three structural shifts illustrate the magnitude of permanent change:

Regulation NMS (2005, effective 2007) introduced the Order Protection Rule, fragmenting US equity trading across dozens of venues. Strategies that relied on specific exchange microstructure — particular order book behaviors, latency patterns, or routing inefficiencies — found their assumptions no longer valid.

HFT proliferation between 2007 and 2013 changed the nature of liquidity. Quoted depth became less reliable as a signal because HFT market makers withdrew quickly in response to order flow. Statistical arbitrage strategies that modeled quote dynamics based on pre-HFT data were trading a different instrument than they thought.

Options market growth has altered how large participants hedge equity exposure. Dealer gamma hedging creates mechanical buying and selling flows that are largely absent from historical data prior to the current options environment. Strategies that model equity price behavior without accounting for these flows are missing a structural feature of the current market.

When structural change is the cause of decay, there is no waiting it out. The market that generated the backtest no longer exists.

Detecting Decay in Real-Time

The statistical framework for decay detection centers on a simple question: is the strategy's current performance consistent with what the backtest predicted for this period?

Rolling Sharpe ratio is the primary signal. Calculate Sharpe over a rolling 90-day window and over a rolling 180-day window. Plot both against the historical distribution from the backtest. When the rolling Sharpe falls below one standard deviation of the expected backtest distribution for the same number of observations, that is worth investigating. When it falls below two standard deviations, it is a strong signal of structural change rather than variance.

Performance vs. backtest expectation is a more direct test. If your backtest showed an average trade of $X with a given win rate, and your live trading shows average trades and win rates that have diverged significantly from those parameters, the strategy is no longer executing in the environment it was designed for. This requires meticulous trade-level data from both backtest and live trading.

Regime mismatch detection compares the current regime signal (ADX, rolling volatility, trend strength) against the regime distribution in the backtest period. If the backtest ran during a period of elevated trending and the current market is in sustained mean-reversion, the strategy's performance weakness may be regime-explained rather than edge-explained. This distinction matters for the response: wait for regime rotation versus retire the strategy.

See Backtesting Is Not Prediction for a deeper treatment of how backtest statistics relate to forward performance expectations.

The Oyamori Approach

Oyamori treats edges as perishable assets. Each strategy in the catalog carries continuous performance monitoring: rolling Sharpe calculations, regime mismatch flags, and alerts when live behavior diverges from backtest expectations by a defined threshold.

When a strategy enters a monitoring window — rolling Sharpe below threshold for a sustained period — the system does not automatically retire it. It escalates to review. The review process distinguishes between regime-explained underperformance (temporary, no action required) and crowding or structural decay (retire and rotate).

The catalog model is built around this rotation. Strategies are researched, validated, deployed, monitored, and retired. The pipeline of new strategy research is treated as infrastructure — not optional maintenance, but the operational core of a systematic trading program. A catalog with no incoming research has no mechanism to replace edges that decay. That is not a trading program. That is a countdown.

Next: Overfitting vs. Robust Strategies →