Why can't you tell from the equity curve alone?
An equity curve cannot separate decay from variance because a strategy with a real, unchanged edge produces losing runs that look exactly like the early part of a strategy that has failed.
The size of those runs is larger than intuition suggests. A rule with a forty percent win rate will, over a few hundred trades, produce a run of eight or nine consecutive losses as a matter of course. A rule with a positive expectancy will spend long stretches below its previous high. None of that is evidence of anything; it is what the distribution of outcomes looks like when it is sampled.
The difficulty is that a strategy which has genuinely stopped working produces the same picture at first. The first ten trades after an edge disappears are not distinguishable from the first ten trades of a normal drawdown. Only the accumulation of evidence over more trades separates them, and by then the question has usually been decided emotionally.
So the curve is the wrong instrument. It shows what happened; it does not show what was expected, and the comparison between the two is the whole question.
How do you set the line before you trade?
Set the line by taking the backtest's own distribution of outcomes, working out the losing streak and drawdown it says are plausible at the strategy's trade count, and writing down the level beyond which the live record would be outside that distribution.
The inputs are the backtest's win rate, its distribution of R-multiples and its trade frequency. From those, the longest losing streak to expect over a given number of trades follows directly, and a resampling of the backtest's trades gives the range of drawdowns the same rule could produce by chance. The line is drawn at the edge of that range, at whatever confidence level you can live with.
Writing it down is not a formality. A line set after the drawdown has begun will be set wherever the current pain is, which makes it worthless. A line set beforehand is a commitment made by the version of you who was thinking clearly, to be honoured by the version who is not.
The line has two parts. One is a maximum drawdown in R or in percent; the other is a maximum losing streak or a minimum rolling expectancy over the last N trades. Either being crossed is a signal; both being crossed is a stronger one.
Which signals separate decay from variance?
Beyond the drawdown itself, four trade-level signals separate decay from variance: a change in how far winning trades travel, a change in how often the setup appears, a shift in results by market condition, and a change in fill quality.
The travel of winners is the earliest. If the rule's winning trades used to reach two R and now stall at one before reversing, the structure the rule exploits may have changed even if the win rate has not moved yet. The maximum favourable excursion of each trade, compared with the backtest's distribution of the same figure, shows this before the equity curve does.
Setup frequency is the second. A rule that fired twice a week and now fires twice a month is seeing a different market, and its results on the trades it does take may or may not carry over. A frequency change is not decay in itself, but it is a reason to look.
Results by condition and fill quality close the list. A strategy that only ever made money in one regime, and is now in another, has not decayed; it was always conditional, and the condition has changed. And a strategy whose entries are being filled worse than the backtest assumed is losing to costs rather than to the market, which is a different problem with a different fix.
What should you do when the line is crossed?
When the line is crossed, stop trading the rule, re-grade it on the most recent history as an out-of-sample test, and resist the instinct to re-optimise it on the period that just hurt.
Stopping is the part that has to be automatic, because the line exists so that the decision does not have to be made under pressure. Continuing past a pre-committed limit because this time feels different is the behaviour the limit was designed to prevent.
Re-grading is the diagnostic. Run the unchanged rule on the recent period as if it were held-back data. If the result is within the backtest's distribution, the line was set too tight or the sample was unlucky, and the rule can be resumed with the line reconsidered. If the recent period is well outside the distribution, the evidence of decay is real.
Re-optimising on the losing period is the trap. It will produce a version of the rule that would have done well on the period that just ended, which is a fit to that period and nothing more. A rule that needs re-tuning after every drawdown has no stable edge to re-tune; it has a series of fits.
How does Strategy Health in QuantParadox handle this?
Strategy Health in QuantParadox compares a strategy's paper or live trade record with the distribution its backtest produced and raises an alert when the recent record falls outside the band, which turns the decision into a comparison made in advance rather than a judgement made in a drawdown.
The comparison is on the measures above rather than on the equity curve alone: rolling expectancy, losing streak, drawdown, and the excursion of trades relative to what the backtest saw. An alert says that the record is outside the range the backtest considered plausible; it does not say why.
That is the limitation to be clear about. The module needs the trade record, imported or produced through the platform's own paper trading, and it cannot distinguish a market that has changed from an execution that has degraded without the fill data to compare. An alert is a reason to re-grade, not a verdict that the edge is gone.
The re-grade itself is the standard out-of-sample test on the recent period, which is the same procedure the rest of the platform uses. A strategy that was never established in the first place, because its trade count was too small, has no band to fall outside of, and Strategy Health says so rather than inventing one.