What Is Edge Decay and How It Shows Up in a Journal

Every strategy that earned a positive expectancy on real fills will, at some point, stop earning that same expectancy. The question is not whether the edge fades. The question is whether the trader notices in the journal data before the equity curve forces the conversation. This post defines what edge decay actually is, walks through the journal signals that show up first, and names the conventions for telling decay apart from a normal losing streak.

By Imperial Analytics

What edge decay actually is

Two distinctions to hold from the start. Edge decay is not the same as a losing streak. A losing streak is a run of negative outcomes around an unchanged distribution. Edge decay is a change in the distribution itself. The strategy's win rate, average winner, average loser, or some combination of the three has shifted. The equity curve eventually reflects that shift, but the per-trade expectancy reflects it sooner.

Edge decay is also not the same as a bad week. A bad week is a sample-size problem. Five trades in a row that close at a loss can happen inside a perfectly stable distribution with a sixty percent win rate. That is variance, not decay. The distinction matters because the right response to variance is to keep executing the plan, while the right response to decay is to stop trading the setup until it has been re-tested.

Why edges decay in the first place

The market-regime cause is the one most often discussed. A breakout setup that worked through a low-volatility expansion phase can stop working when realized volatility doubles and breakouts become traps for fast reversal. The price action that defined the setup looks the same on the chart. The follow-through that paid the trader is gone. A trader running the setup unchanged through both regimes will see the journal record more entries that hit the stop and fewer that reach the target.

The participant-behavior cause is more difficult to measure but no less real. A pattern that paid the trader for two years may have been an artifact of how a class of participants positioned around a particular event. When that class changes its behavior, the pattern decays. The trader cannot see the participant change directly. They can only see the trade outcomes change.

The execution cause is the one most often missed. A trader who learned the setup at one size may slip into different fills, different stop placements, or different exit discipline as the strategy moves to live size. The setup's theoretical edge is unchanged. The realized edge has decayed because the trader executes it differently. The journal is the only place this distinction shows up.

How edge decay shows up in a journal before P&L

The first metric to watch is the ratio of average winner to average loser, often called reward-to-risk. A strategy that ran at a clean two-to-one ratio for a year and now runs at one-point-three-to-one has a real signal worth investigating. Total P&L may still be green because a single outlier winner can lift the average dollar number for many sessions. The ratio is the cleaner read because it normalizes the math.

The second metric is win rate inside the matching condition. The matching condition matters because aggregate win rate is too coarse. If the strategy is defined by a specific setup with a specific trigger, the question is whether trades that pass the setup filter still win at the historical rate. A drift from a sixty percent win rate on filtered trades to a fifty-two percent win rate on filtered trades, holding the average winner and loser steady, is enough to move expectancy from clearly positive to roughly breakeven.

The third metric is maximum favorable excursion per trade. This one shows up earliest of all in some strategies. If the average trade used to reach a peak of one point five times the trader's stop distance before reversing and ending at the target, and the new average peaks at zero point seven times the stop distance, the trades are no longer behaving as designed. The trader may still exit at small wins or breakeven because their exits absorb the change. The MFE record exposes it.

A fourth, slower signal is the distribution of trade durations. A momentum setup that used to deliver winners in twenty minutes and now delivers winners in two hours may be paying the same total dollar amount with a higher cost of capital and a higher cognitive load. That is not decay in the per-trade math, but it is decay in the strategy's practical viability for the trader.

How to tell edge decay from a normal losing streak

The practical version of the test is straightforward. Pick a window large enough to clear the sample-size minimum for the strategy. Twenty trades in the matching condition is the floor used by the AI Operating Charter for any behavioral pattern claim, and the same floor is reasonable for a strategy claim where the trade count is small. Compute win rate, average winner, and average loser for that window. Compare those three numbers against the prior comparable window. If the change in win rate is small and the change in average winner and average loser is small, the distribution is roughly stable and a current drawdown is variance. If the change is large in any of the three, the distribution has shifted and the right read is decay.

The arithmetic of variance helps frame "small" versus "large." A strategy with a sixty percent historical win rate measured over four hundred trades has a standard error around that win rate of roughly two and a half percentage points. A new window of twenty trades has a much larger standard error, roughly eleven percentage points. A new-window win rate of fifty percent inside a historically sixty percent strategy is not, by itself, conclusive evidence of decay; it is well within the normal sampling band for twenty trades. A new-window win rate of thirty-five percent is harder to dismiss. The smaller the new sample, the larger the shift has to be before the decay read is honest.²

The other half of the test is to look at the conditions, not just the outcomes. If the strategy's setup is well-defined, the trader can ask whether the conditions producing the trades have changed. Has realized volatility on the instrument doubled or halved over the window? Has the time-of-day distribution of triggers shifted? Has the average trade duration moved? Conditional changes that line up with outcome changes raise the probability that the read is decay rather than streak. Outcome changes without conditional changes are more consistent with variance.

What to do once edge decay is suspected

The reason for forward-testing at small size rather than backtesting is that the regime change is in the current market, not in the historical record. A backtest will confirm what the journal already shows, which is that the historical record favored the setup. The current question is whether the setup still pays in this regime. Only live data on smaller size can answer that, and the smaller size protects the account while the answer accumulates.

The structural protection against being slow to notice is the journal itself. A trader who tracks expectancy in rolling windows of twenty, fifty, and one hundred trades, and who tags each trade with the setup, instrument, and conditions, will see decay show up in the twenty-trade window long before it shows up in equity. The cost of running the journal at that detail is the cost of the discipline to record it. The reward is the difference between recognizing edge decay at the per-trade level and recognizing it when the account has already given back the cushion the prior edge built.

Frequently asked questions

Frequently asked questions

• q: How many trades does it take before edge decay can be claimed honestly? a: The honest minimum is roughly twenty trades in the matching condition, and even that is a rough floor. A twenty-trade sample carries a wide standard error around the win rate, so the per-trade metrics have to shift by more than that error band before the decay read clears variance. Strategies with smaller per-trade sample sizes need more trades before the call can be made with confidence.
• q: Is edge decay the same as overfitting? a: They are related but not identical. Overfitting is when a strategy's apparent edge was an artifact of fitting noise in historical data and never existed in real time. Edge decay is when a strategy did have a real edge on live trades and then lost it. Both produce a falling expectancy on new trades; the distinction is whether there was ever a real edge to lose.
• q: Can a decayed strategy come back? a: A regime-driven decay can reverse if the regime that supported the original edge returns. A participant-behavior decay or an execution-drift decay is less likely to reverse on its own. The honest move is to treat the decay as the new baseline until the data forces a different read. Hoping for a regime to return is not a strategy decision; it is a wait.
• q: What is the difference between edge decay and a drawdown? a: A drawdown is a fall in equity from a prior peak. It can happen inside an unchanged distribution, which is variance, or it can happen because the distribution has shifted, which is decay. The drawdown describes the equity path; edge decay describes a change in the per-trade math underneath it. Telling them apart is what the per-trade metrics are for.

Sources

Footnotes

1. Aronson, David. Evidence-Based Technical Analysis: Applying the Scientific Method and Statistical Inference to Trading Signals. Wiley, 2007. Establishes the random-walk null hypothesis as the baseline against which any claimed edge must be measured, and the statistical-inference framework for distinguishing real edges from data-mined patterns. ↩

2. The standard error on a sample win rate is approximately the square root of p(1-p)/n, where p is the historical win rate and n is the sample size. For p = 0.60 and n = 20, the standard error is roughly eleven percentage points. For n = 400, the standard error is roughly two and a half percentage points. The formula is the standard one for a binomial proportion and is covered in any introductory statistics text. ↩