Do moving-average crossovers hold up after costs?

01Question

Moving-average crossovers are among the first rules most traders write. Buy when a fast average crosses above a slow one, sell when it crosses back. They are easy to state and easy to backtest, which also makes it easy to find a setting that looks good on one chart. We ask a narrow question across many charts at once: did a standard crossover time its trades better than chance, and did it still make money once realistic costs were paid?

Those are two different questions, and the paper keeps them apart. Costs decide whether a rule made money. A baseline of random trades decides whether its timing added anything. A rule can pass one and fail the other.

02The rule and the baseline

The rule is the one a SixOne8 user would build: a long strategy that enters when EMA(20) crosses above EMA(50) at a bar's close, and a short strategy that enters on the opposite cross. Each position closes on the opposite cross or at a protective stop three ATRs from the entry, whichever comes first. Every market in the catalogue with history from October 2022 is included: 37 currency pairs, 4 metals and 40 crypto markets. The first year warms the averages up; trades are counted from October 2023.

For each market and timeframe we then draw 1,000 sets of random trades with the same number of trades, the same mix of long and short, and the same holding periods, placed at random without overlapping. A random set holds the same exposure to the market's trend as the rule and pays the same costs. The only thing it lacks is the rule's timing. Excess is the rule's annual return minus the median random set's.

03Results

Figure 1 follows each asset class on 4-hour bars. Crypto rose well above its baseline from the first months and stayed there. Forex drifted below zero and below its own baseline. Metals, only 4 markets, swung widely and finished a few points below their baseline.

The crypto line is lumpy. 67% of its final gain was booked in one month, December 2024, as positions opened during the November rally closed. A trend rule earns most of its return in a few long trends and gives some back between them, so a different three years, with fewer or shorter trends, would give a different line.

Table 1 gives the figures for every class and timeframe. The pattern holds across all three timeframes. Forex excess was −3.0 pts, −1.8 pts and −1.8 pts a year on 1h, 4h and 1d, each interval below zero. Crypto excess was +32 pts, +42 pts and +33 pts, each interval above zero.

Metals show why the baseline matters. On daily bars the median metal returned +17.3% a year after costs, which looks like a success. Its random baseline returned almost the same, and the excess was +0.3 pts. Metals rose over the period, and any rule that was long often enough made money. The crossover's timing added nothing to that.

04Across markets

Pooling all 81 markets gives a median 4h excess of +0.2 pts, a number that describes no market at all. Figure 2 shows why: a cluster of currency pairs just below zero and a long tail of crypto markets far above it.

Figure 3 shows the same thing one market at a time, with each rule beside its own baseline. Read down the colours of the names. A handful of coins sit at the very top, where the rule did worst of all; below them the currency pairs gather, short of their baselines by a few points; and most of the coins fill the bottom, clear of theirs by a wide margin. Crypto holds both the best and the worst results, which is what a large effect in a volatile class looks like.

We then asked whether any single market was distinguishable from its own baseline once we allow for having tested 243 market-and-timeframe combinations. Ranking each against its 1,000 random sets and controlling the false-discovery rate at 10% (Benjamini–Hochberg), 0 passed. One market's three years of trades is too few to tell timing from luck. The crypto result rests on the class as a whole, and because coins move together, 40 crypto markets are much less than 40 independent confirmations.

05Parameter stability

A result that appears at one setting and vanishes at its neighbours is usually noise. We repeated the test for 24 pairs of fast and slow averages. Figures 4 and 5 show the median excess for crypto and for forex on 4-hour bars.

Neither picture has an isolated peak. Every crypto setting was above its baseline and every forex setting below. The headline numbers use the 20/50 default chosen before the study, not the strongest cell, which on crypto would have looked considerably better.

06Sensitivity to costs

Faster bars mean more trades, and every trade pays twice. On 1-hour crypto the rule traded about 169 times a year per market. At the 12 basis points per side SixOne8 suggests for crypto, costs came to about 40 points a year. That is more than the gross result of +34, leaving −6.0% net. Figure 6 shows where each timeframe crosses zero as costs rise.

Forex costs are far lower, 1.5 pips or about one basis point per side on most pairs, and still the rule lost money on every timeframe, because it lost before costs. Table 2 sets out what each version could carry.

07Methodology

Every figure in this paper comes from the settings below. Where we made a choice, we state it.

Data

1-hour candles built from 1-minute Twelve Data candles stored by SixOne8, 6 October 2022 to 30 September 2026. 4-hour and daily bars are folded from them on the app's own grid: New York 17:00 for forex and metals, UTC for crypto.

Universe

Every forex, metals and crypto market in the SixOne8 catalogue with history from October 2022: 37 currency pairs, 4 metals and 40 crypto markets. No market was removed after the fact.

Window

Trades from 1 October 2023 to 30 September 2026. The year before warms the averages up, so every setting starts from fully formed values.

Entry

EMA(fast) crosses above EMA(slow) at a bar's close for a long, below for a short; the app's crossover operator, with SMA-seeded EMAs as the app computes them. Filled at the next bar's open. Default 20 and 50, chosen before the study.

Exit

The opposite cross, filled at the next bar's open, or a stop 3 × ATR(14) from the entry, filled at the stop or at the open of a bar that gaps through it. No target and no time limit.

Costs

SixOne8's suggested spread per side for each market: 1.5 pips on currency pairs, 2–8 bps on metals, 12 bps on crypto. Sensitivity from 0 to 20 bps in Figure 6.

Baseline

For each market and timeframe, 1,000 random sets (200 for the parameter grid) with the rule's trade count, sides and holding periods, shuffled and placed without overlap.

Statistics

Per-market rank against its baseline, Benjamini–Hochberg at 10% across all 243 combinations. Intervals for medians from 2,000 bootstrap resamples of markets.

Checked against

The SixOne8 backtest engine, running the same two strategies on EUR/AUD, BTC/USD and XAU/USD (4h, same window). BTC/USD gave the same 118 trades. On EUR/AUD 82 of the engine's 84 trades matched; the other two entered at a Sunday reopen four minutes later. On gold all 89 of the engine's trades matched, and this simulation took 11 more at the reopen after gold's daily break, where a SixOne8 backtest currently takes no trade.

08Limitations

Three years is short, and for crypto it was one strong cycle: long trends up and sharp falls, which is the market a trend rule suits. A different three years could look different. The crypto universe is the coins still in our catalogue with four years of history, so coins that collapsed or were delisted before today are missing, and their absence flatters any long-biased result. Costs are a fixed spread; real costs widen in fast markets, when a crossover is most likely to trade.

The study describes how the rule behaved. It says nothing about whether it will behave the same way next.

09Test it yourself