Equity-Curve Trading: Turning a Strategy On and Off with an Account Moving Average | NOONOO TRADING
Pause when a losing streak begins and resume when the account curve recovers. It looks like a rule that skips losses and often improves a backtest. Yet results frequently worsen when it is used in live trading. This article uses calculations to distinguish when such a rule can and cannot work.
What is equity-curve trading?
Equity-curve trading switches a strategy on or off based on the account curve rather than the market's trade signals. As with a moving average on a price chart, a moving average is applied to cumulative PnL. Live trading stops below the average and resumes above it.
Variants differ, but all have pause and resume conditions. Signals continue to be tracked on paper during the pause so the curve keeps updating. If recording stops, there is no curve on which to evaluate resumption.
Account: $10,000. Per-trade PnL: ±$200, or 2%.
Pause: cumulative PnL curve < its last-20-trade moving average.
Resume: cumulative PnL curve > that moving average.
During pauses, signals are recorded only and the curve updates.
→ No real fills, but the curve continues.
The attraction is clear. If it could selectively avoid losing streaks, maximum drawdown would fall, allowing the same strategy to run at greater size. The crucial assumption is whether it can actually select those periods.
What must be true for the rule to work?
For the curve to predict subsequent results, trade outcomes must have autocorrelation. Pausing after a loss is beneficial in this argument when another loss is more likely than a win.
One check assigns +1 to wins and −1 to losses and calculates correlation between successive trades, or lag-1 autocorrelation. The source uses an approximate chance range of ±2 ÷ √n.
Chance range = 2 ÷ √200 = ±0.14.
Measured lag-1 autocorrelation −0.03 → inside the range → no detected structure.
Measured lag-1 autocorrelation +0.22 → outside the range → worth investigating.
The source says most individual strategies fall into the first group:
a loss does not make the next loss more likely.
Near-zero autocorrelation is interpreted here as the curve's shape carrying no information about the next trade. Under that interpretation, filtering its moving average resembles predicting a coin toss from a sequence of heads. Losing streaks occur naturally; losing-streak probability explains how to calculate typical ranges.
Shuffle checks: why a backtest can look better
The source argues that equity-curve filters almost always improve a fitted backtest because choosing a moving-average period from the historical curve selects a period that avoids its losing stretches. This is a textbook form of overfitting.
Its simplest proposed check is to shuffle trade order. Keep the same outcome set but randomize its sequence, preserving the aggregate outcomes while changing the curve's shape. The guide argues that a genuine filter benefit should remain meaningful under this comparison.
Unfiltered net PnL: +$1,000.
Filtered net PnL: +$1,400, apparently better.
Apply the same filter after 1,000 order shuffles of the same 200 trades:
Average: +$980.
5th–95th percentile range: +$300 to +$1,700.
The original +$1,400 is inside that range.
→ The source attributes it to luck in trade order, rather than demonstrated filter effectiveness.
The guide's adoption rule requires the original result to exceed the shuffled distribution's top-5% threshold. Monte Carlo simulation uses similar shuffling to examine adverse drawdowns. See backtesting for reasons not to accept historical results at face value.
Switching off and on has a cost
A less visible cost is that the resume condition is met only after the curve has already recovered. The first part of a recovery is therefore missed by construction.
Pause triggers among 200 trades: 12.
Average pause: six trades.
→ Skipped trades = 12 × 6 = 72.
Combined PnL of those 72 trades: +$800.
Wins were concentrated just after losing streaks.
Filter cost: −$800.
Re-entry slippage and fees are additional.
The original article states that this value averages near zero for a zero-autocorrelation strategy, while variance increases substantially. Fewer trades mean a smaller sample and results more dependent on luck. This is its explanation for a filter sometimes producing more variability rather than stability.
A further practical problem is recordkeeping. The curve requires continued records during a pause, but people may become less diligent when they are not trading live. After several days, missing data can turn resumption into a discretionary guess. Without an automated trading journal, maintaining the rule itself can be difficult.
Reducing size instead of switching off
On/off is an extreme choice. The source proposes adjusting position size according to drawdown as a way to pursue the same purpose with fewer side effects.
If the next 15 trades would produce +$1,200:
Stop completely → $0, missing the entire recovery.
Half size → +$600.
If those 15 trades would produce −$1,200:
Stop completely → $0.
Half size → −$600.
Half size halves both sides.
A pause blocks recovery as well as losses.
Moreover, fixed-fraction risk automatically reduces size. Risking 2% of the account means order quantity falls as the account shrinks, without an extra drawdown rule. See the position-size calculator and portfolio heat for sizing and combined risk across positions.
$10,000 account → $200 risk per trade.
After 20% drawdown, $8,000 account → $160 risk.
After 40% drawdown, $6,000 account → $120 risk.
Larger losses lead to smaller bets,
lengthening the path toward a zero balance in this model.
The source says this property substantially lowers risk of ruin. The Kelly criterion provides a formula-based approach to an optimal fraction, though the guide describes using half or less of the calculated value as common in practice.
When an equity-curve filter may still help
The method is not useless in every case. The source identifies two situations with a rationale.
1. Regime-dependent strategies. A trend-following strategy can lose repeatedly in a range lasting days or weeks. Such losses can indicate a change in market conditions rather than random sequencing, producing detectable autocorrelation. Even then, directly observing volatility or trend strength can respond faster than the equity curve.
2. Detecting a broken strategy. Change the purpose from avoiding losses to deciding whether to retire a strategy. Below a defined curve threshold, revalidate the strategy and retire it if it fails. The stop threshold must be written down in advance.
Backtest maximum drawdown: 18%.
Live stop threshold: 18% × 1.5 = 27%.
At 27% → stop live trading and revalidate.
Failed revalidation → retire the strategy.
Set the threshold before reaching it.
The source warns that setting it afterward invites self-serving choices.
Live drawdown commonly exceeds backtest maximum drawdown. The source treats approximately 1.5 times the historical maximum as a reason to suspect the strategy no longer behaves as before. See drawdown for interpreting these measures.
Five practical checks
1. Measure lag-1 autocorrelation before adding a filter. The source treats a value inside ±2 ÷ √n as no detected information in the curve.
2. Do not adopt a filter that fails the shuffle check. Backtest improvement alone is not evidence.
3. Count the moving-average periods tested. Selecting the best of 10, 20 and 50 requires a stricter evaluation.
4. Consider size adjustment before on/off switching. Pauses block recovery as well as losses.
5. Automatically record trades during pauses. Missing records turn the resume rule into guesswork.
Three key points
1. The source's condition for an equity-curve filter is autocorrelation in trade outcomes. It says most strategies fall within ±2 ÷ √n, with no detected structure.
2. Backtest improvement is not proof. Its criterion is a result above the top-5% threshold of the shuffled comparison.
3. Pausing blocks recovery periods as well as losses. Fixed-fraction risk is presented as an alternative that reduces size automatically.
Caution
The account sizes, trade counts, PnL amounts, autocorrelations and shuffled distributions above are hypothetical examples illustrating calculations, not measured performance of a specific strategy. The approximate lag-1 chance range ±2 ÷ √n assumes independent outcomes from the same distribution. In real markets where win probabilities vary over time, treat it only as a reference. Leveraged trading can lose all principal. Investment decisions and responsibility remain your own.
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 →