Detecting Decay in a Live Strategy

Every strategy has losing weeks and losing months even while working perfectly as designed — that's what it means for returns to have volatility rather than being a straight line. The hard problem isn't spotting a bad month, it's telling the difference between ordinary noise and the early stages of genuine decay, where the underlying edge is quietly weakening and the bad stretch is the first visible sign of a longer downward trend rather than a random dip that will revert.

What to watch besides P&L

P&L alone is a poor early-warning signal because it's the noisiest thing a strategy produces — by the time cumulative losses are obvious, a strategy has often been decaying for a while already. More reliable early indicators sit upstream of P&L, closer to the mechanics of why the strategy makes money in the first place:

Indicator	What decay looks like
Signal strength (e.g. information coefficient)	A sustained downward drift in a rolling measure of predictive power, not just a single bad period
Turnover	Rising turnover for the same target exposure, often a sign the signal is chasing a weakening effect harder to compensate
Realised transaction costs	Costs creeping up relative to backtested assumptions, eating into an edge that hasn't actually changed on paper
Hit rate / win rate	A gradual decline in the fraction of profitable positions or periods, even if average position size is unchanged
Capacity utilisation	The strategy's assets under management approaching, or exceeding, what the market can absorb without excess impact
Correlation to known crowded factors	Rising correlation to a widely-known factor or strategy, suggesting the edge has become common knowledge

The common thread across these is that they measure the mechanism behind the strategy's returns rather than the returns themselves, and mechanisms tend to move more slowly and more informatively than the noisy P&L series built on top of them.

Worked example

A quantitative equity strategy has run cleanly for three years with a rolling six-month information coefficient hovering around 0.04. Over the following two quarters, the rolling IC drifts down to 0.025, then 0.018, while cumulative P&L over the same window is roughly flat — not yet a visible drawdown, because position sizing has been mechanically increasing to compensate for the weaker signal, keeping expected returns similar even as the underlying edge shrinks. A team watching only the P&L line sees nothing alarming. A team watching the rolling IC sees a signal that has lost more than half its strength in six months and flags it for review well before the compensating position-size increase eventually runs out of room and losses show up directly. The team's investigation finds turnover has also crept up over the same period, consistent with the strategy working harder for a smaller edge — a pattern worth distinguishing from a sudden, discrete break, which points toward a different set of causes covered in Diagnosing a Slump: Decay, Crowding or Costs.

In practice

• Build a rolling dashboard of mechanism indicators, not just cumulative return. Signal strength, turnover, and realised-versus-backtested cost are the three most informative series to track continuously.
• Set a decay threshold in advance, tying back to the discipline of Defining Kill Criteria Before Launch — a rolling IC floor or turnover ceiling agreed on before performance starts sliding is far more useful than a judgment call made under pressure.
• Distinguish gradual decay from a discrete structural break. A signal weakening steadily over quarters points toward crowding or a fading effect; a signal that breaks suddenly points toward a regime change or a data or execution problem, a distinction covered further in Diagnosing a Slump: Decay, Crowding or Costs.
• Check whether the effect is becoming common knowledge. Rising correlation to a well-known factor, or a strategy's returns starting to move with a widely tracked style, is one of the more reliable tells that an edge has been arbitraged down by Factor Crowding.
• Don't wait for statistical certainty before acting. Decay detection is inherently a judgment call made on incomplete evidence — the goal is catching a plausible pattern early, not proving it beyond doubt before responding.