Has a Strategy Decayed, or Is It in a Slump?
A strategy that tested well and operated successfully for months starts struggling. Continuing and stopping can both be wrong: stopping a viable strategy can miss recovery, while continuing a decayed one can drain capital. This guide describes a numerical framework for distinguishing a slump from deterioration.
A slump can be part of normal operation
A profitable strategy still has losing periods. Recent disappointment alone cannot determine whether it has failed. The question is whether the observed results fit its expected variability.
Record two baseline values first: per-trade expectancy and standard deviation. See expectancy and R multiples. Without a baseline, later judgments lack a comparison. Much of monitoring begins before live operation.
Small samples cannot reliably detect decay
Suppose expected return is +0.25R per trade and standard deviation is 1.2R. First examine how far a sample mean can move while those assumptions remain unchanged.
μ = +0.25R; σ = 1.2R.
Standard error SE = σ ÷ √n.
50 trades
SE = 1.2 ÷ 7.07 = 0.1697R.
0.25 ± 1.96 × 0.1697
= −0.083R to +0.583R.
100 trades
SE = 1.2 ÷ 10 = 0.120R.
Range: +0.015R to +0.485R.
200 trades
SE = 1.2 ÷ 14.14 = 0.0849R.
Range: +0.084R to +0.416R.
These normal-approximation ranges assume the stated stable distribution and appropriate independence.
A 50-trade mean of −0.08R remains inside this illustrative range. By contrast, +0.05R across 200 trades lies below its lower boundary. Interpretation depends on sample size, not just the recent average.
The guide therefore warns against declaring a strategy dead from 30–50 trades alone. Such a sample can have insufficient power to distinguish deterioration from noise. See trade sample size.
Two tests ask different questions
Suppose the latest 100 trades average −0.05R with a 1.3R standard deviation. Two null hypotheses produce different answers.
t = (−0.05 − 0.25) ÷ (1.3 ÷ √100)
= −0.30 ÷ 0.13 = −2.31.
Two-sided p ≈ 0.023.
Reject the fixed-baseline null at 5% under the example's assumptions.
Question 2: Is the present mean different from zero?
t = (−0.05 − 0) ÷ 0.13 = −0.38.
Two-sided p ≈ 0.70.
Do not reject zero mean.
This does not establish a negative expected return.
The sample can support deterioration from the old baseline without establishing that current expectancy is negative. The guide proposes reducing exposure and gathering evidence as one research-policy response, rather than treating the two questions as identical. Position sizing and the statistical assessment are separate decisions.
Cumulative deviations: Looking for when the path changed
A rolling mean requires a window choice. The guide's simple cumulative-deviation chart instead adds each realized R minus the baseline mean from a defined starting point.
S(n) = Σ[r(i) − 0.25].
Under the unchanged independent model, S behaves like a zero-mean random walk.
Its standard-deviation scale is σ√n.
At 100 trades
1.2 × 10 = 12R.
Observed S(100) = −28R.
−28 ÷ 12 = −2.33 standard deviations.
Illustrative review thresholds
|S| ÷ (σ√n) > 2.0: Review.
> 3.0: The guide's example stop threshold.
Repeated monitoring requires its own false-alarm calibration.
The chart can help locate a change in slope. If it rises through trade 300 and then declines, investigate conditions around that point: fees, listings, volatility regimes or rule modifications. A journal needs dated changes for this investigation.
Gross performance can hold up while net performance deteriorates
Sometimes higher execution cost explains weaker results rather than a weaker signal. The distinction matters.
Initially
Average gross return: +0.34% per trade.
Round-trip fees and slippage: 0.10%.
Net: +0.24%.
Later, with worse fills
Gross: +0.34%, unchanged.
Effective round-trip cost: 0.22%.
Net: +0.12%.
─────────────
The gross signal result is unchanged,
while net expectancy is halved.
Changing signal parameters would not directly address that example's cause. Execution method, order type and trading frequency deserve investigation instead. If gross expectancy itself falls from +0.34% to +0.12%, the deterioration is in the gross return process.
Record gross and net separately so the distinction remains measurable. See backtest assumptions for costs.
One historical maximum drawdown is not a definitive boundary
A backtest's largest drawdown is the worst event on one historical path, not a guaranteed future limit. Exceeding it alone does not establish decay.
Historical maximum drawdown: −18%.
Shuffle trade order 1,000 times.
95th percentile of drawdown magnitude: 27%.
99th percentile: 34%.
Current drawdown −21%
Worse than the historical path but within the simulated 95% magnitude reference.
Not sufficient evidence of decay by itself.
Current drawdown −31%
Beyond the 95% reference, below the 99% reference.
The guide's example policy calls for review and reduced exposure.
See drawdown and losing-streak probability. An eight-loss streak needs the strategy's probability model and sample context before it can be interpreted.
If deterioration is supported, classify its cause
Different explanations call for different research responses.
① No original edge
The backtest was overfit.
The guide favors retiring the candidate rather than repeatedly retuning it.
② Higher costs
Gross holds up; net falls.
Investigate execution.
③ Regime change
For example, a trend strategy enters a range.
Recovery is possible if suitable conditions return; the guide proposes limited exposure during observation.
④ Structural disappearance
Competition removes the exploited inefficiency.
The guide treats confirmed permanent loss of the edge as a retirement case.
① and ④ can be difficult to distinguish in real time. The first often diverges soon after launch, while the second follows a period of stronger operation. Compare the first 100–200 live trades with the prior baseline, and review overfitting and walk-forward validation.
Regime change is particularly uncertain because waiting for recovery can conceal structural decay. Define the observation period and acceptable loss beforehand.
Pre-register review, stop and restart rules
Self-justification can be more difficult than the calculation. Choosing criteria only after deterioration invites favorable reinterpretation. The guide therefore proposes writing them before launching a strategy.
Name, start date and backtest period.
Baseline mean +0.25R; standard deviation 1.2R.
Baseline profit factor 1.62.
Simulated 95th-percentile drawdown magnitude 27%.
Reduce exposure if either occurs:
Rolling 200-trade standardized cumulative deviation below −2.0, or drawdown exceeding 27%.
Example reduction: 50%.
Stop if both occur:
The rolling 200-trade mean's 95% upper bound is below zero, and cumulative deviation is below −3.0.
Then observe for a predefined minimum period before reassessment.
Restart example:
At least 100 paper trades and an expectancy lower bound above zero.
Thresholds require validation appropriate to the data and repeated testing.
Specify restart criteria too; otherwise “it feels better again” can replace the process. See equity-curve trading for another monitoring approach and its tradeoffs.
Five common mistakes
The sample may not distinguish decay from noise.
② Trying many rolling windows and selecting the worst.
Twenty inspected windows create multiple-testing concerns.
③ Retuning parameters after poor results without fresh validation.
This can repeat overfitting.
④ Recording only net outcomes.
Execution cost and gross deterioration become hard to separate.
⑤ Stopping observation when live trading stops.
You then cannot measure whether the strategy subsequently recovered.
Continued paper records after stopping can inform how the monitoring policy performed, rather than relying on memory. See profit factor for interpreting rolling estimates.
Recap
② With μ 0.25R and σ 1.2R, −0.08R across 50 trades fits the illustrated range.
③ Deterioration from baseline, p≈0.023, differs from evidence against zero, p≈0.70.
④ Cumulative deviations help locate changes in slope.
⑤ −28 ÷ (1.2√100) = −2.33.
⑥ Unchanged gross with halved net suggests a cost investigation.
⑦ Historical MDD −18% is not a universal stopping boundary; the example uses a simulated 27% reference.
⑧ Distinguish overfit edge, costs, regimes and structural disappearance.
⑨ Predefine reduction, stop and restart criteria.
⑩ Retuning after deterioration can simply repeat overfitting.
Judge recent results against a predefined distribution and monitoring policy. Without that reference, decisions can follow the urge to stop during losses and resume after recovery. The illustrative numbers explain the method; they are not measured performance or changes to any live trading rules.
NOONOO TRADING invites you to follow live trading in our free chat.
Start in the bot📈 OKX trading fee discount for new registrations
Register for the OKX Fee Discount →