A strategy is easiest to trust immediately after a profitable backtest and hardest to trust after ten consecutive losses. That is precisely when an investor is most likely to replace rules with a story: the market has changed, the robot needs a new filter, the last signal was “obviously” wrong.

The better question is not “is the equity curve currently red?” It is:

The diagnostic question Is this an expected drawdown, or has the data stopped supporting the edge? Answer it with evidence collected before the next trade — not with a post-hoc explanation of the last one.

The same drawdown can mean two different things

In a healthy trend-following system, losses often cluster while the market changes direction repeatedly. The strategy can look inactive or “wrong” for weeks, then recover when a persistent move appears. In a degraded system, the losses may look similar at first — but the recovery never arrives because the signal no longer separates favourable from unfavourable outcomes.

Two drawdown stories The upper curve makes a new high after a temporary drawdown. The lower curve continues making lower highs and has no recovery, which is a stronger reason to investigate. EXPECTED DRAWDOWN / RECOVERY POSSIBLE EDGE DECAY new high time → recovery lower highs, no recovery
Conceptual illustration. A drawdown becomes a model-risk question when the time underwater and the behaviour of new trades move outside the tested range.

Start with the path, not the final return

Two systems can both return 14% over a test period. One may reach the result through frequent small gains and short pauses. The other may spend most of the period flat, lose six times in a row and make its year in two large trends. Their headline returns match; their psychological and operational requirements do not.

For a prop account, the path matters even more. A later recovery cannot undo an earlier daily-loss or maximum-drawdown breach. For a personal account, the same path may be acceptable if the capital and holding period were chosen for it. The system is not judged in isolation: it is judged against the environment in which it must survive.

A concrete example: RegimeBreak 5-20-60

The current PropQuant Lab MT5 research gives a useful example of why a single red number is not enough. In the XAUUSD long test at 0.50% initial risk, RegimeBreak produced 86 trades, a 50.0% win rate, $14,232.03 net profit, 1.86 profit factor and 2.73% maximum equity drawdown. Its worst losing streak was six trades.

Sample86 trades
Win rate50.0%
Profit factor1.86
Max equity DD2.73%
Worst streak6 losses
Average hold33h 54m
RegimeBreak XAUUSD long balance curves at five risk levels
Existing MT5 result: the return curve is shown at five risk levels. Scaling risk changes the distance to the drawdown limit; it does not create a new trading edge.

Now compare the same historical trade sequence at 4.00% risk. Net profit rises to $90,038.59, but maximum equity drawdown rises to 17.03%. Profit factor remains broadly similar at 1.87. The higher return is leverage applied to the same edge, not proof that the strategy became better.

RegimeBreak XAUUSD return and drawdown sensitivity to initial risk
Actual MT5 sensitivity from the published test. The useful comparison is not “which line ends highest?” but “which path can the account tolerate?”
What this example does — and does not — prove

The test demonstrates a stable relationship between risk size and historical drawdown for this sample. It does not prove that the edge will persist, that intraday prop limits cannot be breached, or that 4.00% risk is operationally acceptable. Historical performance is evidence for a decision, not a forecast.

Five signals that deserve a review

1. The drawdown lasts outside its historical range

Depth is intuitive, so it attracts too much attention. Duration is often more informative. If the longest historical period from peak equity to recovery was 90 trading days and the live strategy has been underwater for 180 days, that is a review trigger even if the drawdown percentage is modest.

Do not set a universal “90 days means dead” rule. Estimate the distribution of historical drawdown durations, then compare live time-underwater with a conservative percentile. A trend system may have an ugly quarter and still be healthy; it should not be allowed to drift indefinitely without a defined escalation path.

2. Rolling expectancy turns negative on a meaningful sample

Track the average net result per trade over a rolling window, including commission, spread and swap:

Rolling expectancy E = (win rate × average win) − (loss rate × average loss) − costs Use the same units and cost assumptions as the research test. A window of five trades is a headline; a window of 40–60 trades is more useful.

A negative 20-trade window is not automatically a failure. It may be one bad cluster. A negative rolling expectancy across several consecutive windows, especially when the live distribution is materially worse than the out-of-sample distribution, is a stronger signal.

3. The trade distribution changes, not just the total P&L

Monitor the shape of the results:

  • win rate and average win/loss;
  • loss-streak and holding-time distribution;
  • profit concentration by month, symbol and direction;
  • number of signals and missed/blocked executions;
  • relationship between signal quality and market regime.

For example, if a trend-following system historically earned from a small number of long-running winners, a temporary lack of winners is not surprising. But if winners still occur while their average size collapses, or if losers become larger because exits are no longer filled as tested, the problem may be execution or a changed market microstructure.

4. Live execution drifts away from the research model

A strategy can “break” operationally while its market idea remains valid. Compare the live log with the test on the fields that can actually change the result:

CheckResearch assumptionLive warning
SpreadBroker/tester spread or fixed estimateObserved spread regularly exceeds the reserve
SlippageEntry and exit at a reproducible priceFills move the average trade below expectancy
SymbolContract size, tick value and session are knownBroker specification or trading hours changed
TimingSignal evaluated on the intended barTimezone, rollover or bar-close logic differs
RiskVolume rounded as in the testerMinimum lot or margin changes actual risk

The RegimeBreak research itself states an important limitation: the prop replay uses closed-deal P&L and does not reconstruct intraday open-equity drawdown. That is not a flaw to hide; it is a boundary around the conclusion. A monitoring article should make the same boundaries explicit.

5. Independent validation windows fail together

A backtest can contain a lucky regime. Divide research into development, out-of-sample and forward periods. If one window fails, investigate. If several non-overlapping windows fail under realistic costs, the burden of proof changes: stop searching for a new parameter and ask whether the premise still makes sense.

What is not enough to declare a strategy dead

These events are uncomfortable, but they are weak evidence on their own:

  • one losing month;
  • a six-trade losing streak when six losses were present in research;
  • no trades for several weeks when the rules correctly say “no signal”;
  • one symbol failing with only three trades;
  • a lower return after reducing risk;
  • a strategy underperforming a different strategy in a different regime.

RegimeBreak’s XAUUSD short test is a good cautionary example: it ended with a 0.02 profit factor, but it contained only three trades. That is enough to say “this sample does not support the short side.” It is not enough to estimate a robust long-run failure probability. Small samples create both false confidence and false pessimism.

A practical monitoring protocol

The solution is to decide in advance what gets measured, how often, and what action follows. The thresholds below are an example framework, not a universal law.

StateEvidenceAction
NormalDrawdown and streak remain inside the historical range; execution is in tolerance.Continue at planned risk. Do not optimise.
WatchTime underwater exceeds the historical median or rolling results weaken.Keep collecting data; review costs, timing and regime exposure.
ReduceLive drawdown approaches the stress limit or several rolling windows are negative.Reduce risk and pause new capital until the review is complete.
QuarantineExecution mismatch, code discrepancy or repeated OOS failure.Stop live deployment; reproduce the issue in a controlled test.
ArchiveThe premise fails after realistic costs across independent periods.Retire the version. Preserve the log and write down why.
Strategy review decision flow A five-step flow from observing a drawdown to continuing, reducing risk, quarantining or archiving the strategy. OBSERVEdrawdown COMPAREtested range CHECKexecution + OOS CONTINUE / WATCH REDUCE / QUARANTINE repeated failure → archive
A review should produce an action, not another vague opinion about the latest trade.

The monthly review template

A compact monthly report can contain enough information to prevent emotional decisions:

  1. Equity: current drawdown, time underwater and distance to the previous high.
  2. Trades: count, expectancy, profit factor, win/loss size and streaks.
  3. Execution: spread, slippage, rejected orders, latency and symbol changes.
  4. Regime: market direction, volatility and the conditions under which the strategy is expected to be quiet.
  5. Decision: continue, watch, reduce, quarantine or archive — with a reason.

The most important field is the last one. If the only action is “change the parameters until the curve looks better,” the review has turned into overfitting. A good process allows the system to be wrong and preserves the evidence when it is.

Final rule: do not confuse discomfort with information

A strategy can be profitable and still be unsuitable for a particular account. It can also be in a drawdown and still be behaving exactly as researched. The task is to define the operating envelope before the next uncomfortable period arrives.

For RegimeBreak, that means separating at least four questions: does the signal still produce an edge, does the broker execute it as tested, can the chosen risk survive the account rules, and is the sample large enough to support a conclusion? One red month answers none of them.

Practical rule

Do not change a strategy because it is losing. Change it when pre-defined evidence says its assumptions, execution or distribution have moved outside the tested range.

Conclusion

The professional response to a drawdown is neither blind faith nor immediate optimisation. It is a small, repeatable investigation: measure the path, compare it with the research, check the implementation, and take the action written in the playbook.

That is how an algorithm earns the right to continue — and how a broken version is retired before it consumes more capital.