Tested by BacktestingMax

9/21 EMA crossover backtest: from 5-minute to daily charts

Does the 9/21 EMA crossover make money, and on which chart?

Mostly no. It made money only on some markets, and only on 1-hour, 4-hour and daily charts. The strategy buys when the 9 EMA (a fast average of recent prices) crosses above the 21 EMA (a slower average), and sells when it crosses below. We ran 40 backtests from 2014 to 2025, after trading costs: 8 markets (EUR/USD, GBP/USD, USD/JPY, AUD/USD, gold, the S&P 500, the Nasdaq 100 and the DAX), each on 5 timeframes from 5-minute to daily charts. On 5-minute and 15-minute charts, the crossover lost money on all 8 markets. On 1-hour, 4-hour and daily charts, 10 of 24 tests made money. 5 of them made money in the older years (2014–2021) and again in the newer years (2022–2025): gold on 1-hour and 4-hour charts, USD/JPY on 1-hour charts, the DAX on 4-hour charts and EUR/USD on daily charts.

40 backtests 2014–2025 Trading costs included Tested 12 October 2026

Made money

10

of 40 tests, with the basic rules

Made money in both periods

5

2014–2021 and again 2022–2025

Best backtest

+539%

USD/JPY, 1-hour chart: account growth 2014–2025

With a 200 EMA trend filter

6

tests made money in both periods, compared with 5 with the basic rules

Made money in both periods

• EUR/USD, daily
• USD/JPY, 1-hour
• Gold, 1-hour
• Gold, 4-hour
• DAX, 4-hour

Lost money on every market

• 5-minute charts
• 15-minute charts

The results

Results on every market

Each box shows one backtest of the 9/21 EMA crossover with the basic rules, from 2014 to 2025, after trading costs: the % result, and a small chart of the account balance. The solid line is the account balance and the dashed line is the starting balance. Green results made money; red results lost money. Under each result: whether it made money in both 2014–2021 and 2022–2025, or passed all four of our checks (see "How a test passes" further down).

The account balance

The best test and the middle test

What we found

What the 40 backtests show

1. On 5-minute and 15-minute charts, the 9/21 crossover lost money on all 8 markets. On 5-minute charts, all 8 accounts were down 90% by June 2014, in the first 6 months of the test. Trading costs are the main reason: before costs, the 5-minute trades lost 6.5% of the money they risked in total, and after costs they lost 24.3%. To check that the losses were not only a 2014 problem, we also ran the 5-minute and 15-minute charts on 2022–2025 alone. 15 of the 16 tests still lost money; only USD/JPY on 15-minute charts made money.

2. Gold made money on the most charts: the crossover made money on its 1-hour, 4-hour and daily charts. USD/JPY and the DAX made money on their 1-hour and 4-hour charts. The best backtest was USD/JPY on 1-hour charts: +539% from 2014 to 2025, but at one point the account fell 54% from its highest balance.

3. On EUR/USD daily charts, one sell trade from July 2014 to April 2015 made 24.9R (24.9 times the money it risked). That one trade made more than the whole test made in 131 trades (21.2R), so without it the test would have lost money. With the basic rules, this is the only test that passed all four of our checks (see 'How a test passes'). Across the 40 tests, only 14% to 35% of trades won.

4. On the S&P 500, the crossover lost money on every chart, even though the index rose about 271% from 2014 to 2025. Much of the loss came from the sell trades: on 1-hour, 4-hour and daily charts, the S&P 500 sells lost 257.3R and the buys made 37.7R. In all 24 tests on these three charts, the trades in the same direction as the market's overall move from 2014 to 2025 (buys if it ended higher, sells if it ended lower) did better than the others.

5. We tried 4 changes traders often make: a 200 EMA trend filter, an ADX trend-strength filter, a pullback entry and a 2:1 take-profit. The 200 EMA filter (only buy above the 200 EMA, only sell below it) did best on 2014–2021, the years we used to pick the best change. On 2022–2025, it made only a small difference (see 'Can the 9/21 EMA crossover be improved?' below). It helped most on 4-hour charts: with the filter, 4 of 8 tests made money in both periods, compared with 2 of 8 with the basic rules. No change made the 5-minute or 15-minute charts make money from 2014 to 2025.

Trading costs

Why the 5-minute and 15-minute charts lost

Every trade pays a trading cost: the spread, which is the gap between the buy price and the sell price. Each bar shows that cost as a share of the money a trade risks (what it loses if the stop-loss is hit), for all trades on all 8 markets from 2014 to 2025.

• 5-minute17,748 trades 17.8%
• 15-minute35,089 trades 8.1%
• 1-hour23,859 trades 3.2%
• 4-hour6,245 trades 1.5%
• Daily1,040 trades 0.6%

The stop-loss is 2 times the ATR (how far price moves in an average candle), so on 5-minute and 15-minute charts it is close to the entry price, and the same trading cost is a bigger share of each trade. Most of these 5-minute and 15-minute trades are from 2014 to 2016, because most of those accounts were down 90% by then. On 2022–2025 alone, when prices moved more, the cost was 8.1% of the money risked on 5-minute charts and 4.7% on 15-minute charts. The test does not include overnight (swap) fees, which brokers charge for keeping a trade open overnight.

Buys against sells

On stock indices and gold, the buys made money and the sells lost

We added up all the trades on 1-hour, 4-hour and daily charts from 2014 to 2025, in R. 1R is the money risked on one trade: 1% of the account.

3 stock indices: buys

+455.2R

5,700 trades

3 stock indices: sells

−577.1R

5,653 trades

Gold: buys

+329.1R

1,905 trades

Gold: sells

−90.7R

1,922 trades

The 3 stock indices are the S&P 500, the Nasdaq 100 and the DAX. They and gold all rose a lot from 2014 to 2025. On the 4 currency pairs, the buys lost 31.8R and the sells lost 64.4R. We did not test a version that only buys.

Year by year

Without 2020, the 1-hour, 4-hour and daily tests lost money in total

We added up the 24 tests on 1-hour, 4-hour and daily charts by the year each trade opened, in R (1R is the money risked on one trade).

• 20142,561 trades +62.1R
• 20152,591 trades +43.4R
• 20162,660 trades −270.5R
• 20172,614 trades −72.6R
• 20182,589 trades −3.9R
• 20192,615 trades −95.2R
• 20202,520 trades +310.2R
• 20212,679 trades −107.8R
• 20222,578 trades +137.2R
• 20232,552 trades +26.2R
• 20242,626 trades −27.1R
• 20252,559 trades +18.2R

From 2016 to 2019, the 24 tests lost money four years in a row. 2020 alone made 310.2R, more than all 12 years together (about 20R). The 5-minute and 15-minute tests are left out, because most of those accounts were down 90% by 2016.

The strategy

What the 9/21 EMA crossover is

An EMA (exponential moving average) is an indicator that shows the average price of the last few candles, with more weight on the newest candles. The 9 EMA averages about the last 9 candles, so it follows price closely. The 21 EMA averages about the last 21 candles, so it moves more slowly. After price has been rising for a while, the 9 EMA crosses above the 21 EMA. After price has been falling for a while, the 9 EMA crosses below it. The 9/21 crossover strategy trades in the direction of each cross. Many day traders use it on 5-minute and 15-minute charts.

EMAs

9 and 21, on closing prices

Stop-loss

2 × ATR(14)

Take-profit

None. The next cross in the other direction closes the trade.

Risk per trade

1% of the account

The rules we tested

1. Buy when the 9 EMA crosses above the 21 EMA. We check for a cross when each candle closes, and the trade opens at the next candle's open.
2. Sell (go short) when the 9 EMA crosses below the 21 EMA.
3. When the EMAs cross the other way, close the trade and open a new trade in the new direction. In our test, the new trade opens one candle after the old trade closes. After a stop-loss, wait for the next cross.
4. ATR is an indicator of how far price moves in an average candle, measured over the last 14 candles. Put the stop-loss 2 times the ATR away from the entry price.
5. Size every trade so that, if the stop-loss is hit, the trade loses 1% of the account.

The method

How we tested it

The test

Markets

EUR/USD, GBP/USD, USD/JPY, AUD/USD, Gold, S&P 500, Nasdaq 100 and DAX

Charts

5-minute, 15-minute, 1-hour, 4-hour and daily candles

Dates

January 2014 to December 2025

Account

$10,000, risking 1% of the account on each trade

Trading costs

Included. Every trade pays that market's normal trading cost: the spread, which is the gap between the buy price and the sell price.

How a test passes

We call a test passed only if it meets all four checks. A strategy can make money in one backtest and still fail one of these checks.

• At least 100 trades. With fewer trades, a good result could be luck.
• It made money from 2014 to 2021 and again from 2022 to 2025. This shows whether the strategy kept working in the newer years, not only in the older ones.
• A profit factor (the money won divided by the money lost) of at least 1.2 from 2022 to 2025, after trading costs.
• In 95% of 1,000 random shuffles of the trades, the account must not fall more than 30% from its highest point.

Improvements

Can the 9/21 EMA crossover be improved?

Traders often add one rule to the 9/21 crossover to try to improve it. We tested four common ones, each on the same 40 tests. If we picked the best change and judged it on the same years, it would almost always look good. So we picked the best change using 2014–2021 only, then checked it on 2022–2025, the years we did not use to pick it.

On 2014–2021, the years we used to pick the best change, the 200 EMA trend filter did best: 11 of 40 tests made money with it, compared with 7 with the basic rules. On 2022–2025, the years we kept for the check, the difference was small: 20 of 40 tests made money with the filter, compared with 18 without it. In every version, most 5-minute and 15-minute accounts were already down 90% before 2022, and each trade risks 1% of what is left, so those tests made only small trades, or none, in 2022–2025. The 2022–2025 counts come mainly from the 1-hour, 4-hour and daily charts. On daily charts, the filter left only 58 to 76 trades per test, too few to judge. So the filter helped a little: with it, 6 of 40 tests made money in both periods, compared with 5 with the basic rules.

Passed all four checks with a 200 EMA trend filter

• Gold, 4-hour

Questions

Questions traders ask

Keep exploring

Related strategies we tested

class EmaCrossover921(Strategy):
    """The 9/21 EMA crossover as most traders learn it, plus four changes traders often make.
    Standard (all switches off): buy when the 9 EMA crosses above the 21 EMA, sell short when
    it crosses below. The opposite cross closes the trade and opens a new one the other way
    (the exit fills at the next candle's open, the new trade one candle later). A stop-loss
    2 x ATR(14) away caps each trade; size so that the stop costs 1% of the account. After a
    stop-loss (or a take-profit), wait for the next cross.
    The changes, one at a time:
      trend_filter = 1   only buy above the 200 EMA, only sell below it
      adx_filter   = 1   only open a trade when ADX(14) is above 25 (a trending market)
      pullback     = 1   don't buy the cross itself: after the cross, buy the first candle that
                         dips to the 9 EMA and closes back above it (sells mirror this); the
                         opposite cross closes the trade but doesn't open a new one at once
      exit_rr      = 2   add a take-profit at 2 times the risk; the opposite cross still exits
    """
    params = {"fast": Param(type=int, default=9, min=2, max=100),
              "slow": Param(type=int, default=21, min=3, max=200),
              "atr_mult": Param(type=float, default=2.0, min=0.5, max=5.0),
              "trend_filter": Param(type=int, default=0, min=0, max=1),
              "adx_filter": Param(type=int, default=0, min=0, max=1),
              "adx_min": Param(type=float, default=25.0, min=0.0, max=100.0),
              "pullback": Param(type=int, default=0, min=0, max=1),
              "exit_rr": Param(type=float, default=0.0, min=0.0, max=10.0)}
    lookback = 220
    def init(self, ctx):
        self.fast = ctx.indicator("builtin:ema", period=ctx.params.fast)
        self.slow = ctx.indicator("builtin:ema", period=ctx.params.slow)
        self.atr = ctx.atr(period=14)
        self.ema200 = ctx.ema(source=ctx.close, period=200)
        self.adx = ctx.indicator("builtin:adx", period=14)
        self.reverse_to = None  # "long" / "short": open this once the closing trade is gone
        self.armed = None       # pullback version: the side of the last cross, until it trades
    def on_bar(self, ctx):
        f0, f1 = self.fast.ema[0], self.fast.ema[1]
        s0, s1 = self.slow.ema[0], self.slow.ema[1]
        a = self.atr.value
        if f0 is None or f1 is None or s0 is None or s1 is None or a is None:
            return
        trend = ctx.params.trend_filter == 1
        if trend and self.ema200.value is None:
            return
        use_adx = ctx.params.adx_filter == 1
        adx = self.adx.adx[0]
        if use_adx and adx is None:
            return
        pullback = ctx.params.pullback == 1
        rr = ctx.params.exit_rr
        cross_up = f1 <= s1 and f0 > s0
        cross_down = f1 >= s1 and f0 < s0
        if pullback:
            if cross_up:
                self.armed = "long"
            elif cross_down:
                self.armed = "short"
        pos = ctx.position
        if pos is not None:
            if pos.is_long and cross_down:
                ctx.close_position()
                self.reverse_to = None if pullback else "short"
            elif pos.is_short and cross_up:
                ctx.close_position()
                self.reverse_to = None if pullback else "long"
            return
        entry = ctx.close[0]
        if pullback:
            # The first candle after the cross that dips to the 9 EMA and closes back beyond it.
            # The cross candle itself doesn't count, and the EMAs must still be in cross order.
            go_long = (self.armed == "long" and not cross_up and f0 > s0
                       and ctx.low[0] <= f0 and entry > f0)
            go_short = (self.armed == "short" and not cross_down and f0 < s0
                        and ctx.high[0] >= f0 and entry < f0)
        else:
            # A reversal waits one candle for the close to fill. Drop it if the EMAs crossed back.
            pending, self.reverse_to = self.reverse_to, None
            go_long = cross_up or (pending == "long" and f0 > s0)
            go_short = cross_down or (pending == "short" and f0 < s0)
        long_ok = not trend or entry > self.ema200.value
        short_ok = not trend or entry < self.ema200.value
        strong = not use_adx or adx > ctx.params.adx_min
        if ctx.gate("ema9_crosses_up", go_long) and ctx.gate("long_trend_filter", long_ok) \
                and ctx.gate("adx_strong", strong):
            stop = entry - ctx.params.atr_mult * a
            tp = entry + rr * (entry - stop) if rr > 0 else None
            if ctx.buy(size=ctx.size_by_risk_pct(0.01, stop=stop, entry=entry), sl=stop, tp=tp,
                       tag="ema_long") is not None:
                self.armed = None
        elif ctx.gate("ema9_crosses_down", go_short) and ctx.gate("short_trend_filter", short_ok) \
                and ctx.gate("adx_strong", strong):
            stop = entry + ctx.params.atr_mult * a
            tp = entry - rr * (stop - entry) if rr > 0 else None
            if ctx.sell(size=ctx.size_by_risk_pct(0.01, stop=stop, entry=entry), sl=stop, tp=tp,
                        tag="ema_short") is not None:
                self.armed = None

Your turn

Test the 9/21 EMA crossover yourself

Change the settings, the market or the stop-loss and see the result in seconds. Ask the BacktestingMax AI: Backtest a 9/21 EMA crossover with a 200 EMA filter on gold, 4-hour chart, from 2014 to 2025.

Free account. No credit card needed.

Backtest on historical data with trading costs, run by BacktestingMax. Past results do not guarantee future results, and nothing here is financial advice. Updated 12 October 2026.