ResearchPart 1 of the MT5 Indicator Backtest Series

SMA Crossover Backtest: 25 Variants Tested on EUR/USD 4H (2013–2026)

By Thomas Bayes // BRAHE Research Division·May 2026·12 min read

60-Second Summary

XZero survivors from 25 SMA crossover variants tested
X48% killed by negative expectancy — slow MAs lag so severely entries arrive after the move is spent
X44% killed by IS/OOS drift — win rates learned in-sample do not transfer out-of-sample (mean drift 17.3pp)
+Best performer: SMA 30/200 (+0.862 Sharpe, +23.56 pips net expectancy) — failed drift in all 6 windows
+Sweet spot: fast 25–75 against slow 50 — the only group with consistently positive gross expectancy

Jump to Results Best Performer Parameter Analysis Key Findings FAQ

The simple moving average crossover is the oldest systematic trading signal in widespread use. Whether it translates into a statistically robust edge under modern market conditions and realistic transaction costs is a different question entirely. This article answers it with 13 years of walk-forward validated data across 25 parameter configurations.

1. Methodology

All articles in this series use identical methodology. Results are directly comparable across parts.

Instrument	EUR/USD 4H
Dataset	January 1, 2013 to April 18, 2026 (22,193 bars)
Walk-forward windows	6 non-overlapping (2-year in-sample / 1-year out-of-sample)
Transaction cost	2.0 pip round-trip (1.5 pip spread + 0.5 pip slippage)
Entry	Fast SMA crosses above slow SMA, long at next bar open
Exit	Fixed 8-bar hold (uniform across all SMA variants)
Ranking metric	OOS Sharpe ratio, net of transaction costs

Walk-Forward Windows

Window	In-Sample	Out-of-Sample
1	2013-01-01 to 2015-01-01	2015-01-01 to 2016-01-01
2	2015-01-01 to 2016-12-31	2016-12-31 to 2017-12-31
3	2016-12-31 to 2018-12-31	2018-12-31 to 2019-12-31
4	2018-12-31 to 2020-12-30	2020-12-30 to 2021-12-30
5	2020-12-30 to 2022-12-30	2022-12-30 to 2023-12-30
6	2022-12-30 to 2024-12-29	2024-12-29 to 2025-12-29

Five-Stage Kill Filter

Stage 1

Net Expectancy

Kill if net expectancy is 0 pips or less after transaction costs

Stage 2

Effective Sample Size

Kill if ESS-adjusted OOS trade count is under 50

Stage 3

IS/OOS Drift

Kill if win rate drift exceeds 5pp in any single window — per-window check, not aggregate

Stage 4

Binomial Test

Kill if OOS win rate is not statistically significant (p at or above 0.05)

Stage 5

Sharpe Variance

Kill if Sharpe std dev exceeds Sharpe mean across windows

Stage 3 uses a per-window check, not an aggregate average. A signal with mean drift of 2pp that has one window at 49pp is killed. Aggregate drift averaging conceals window-specific overfitting.

2. Parameter Grid

25 variants tested across 10 fast periods and 4 slow periods. All use an 8-bar fixed exit. The grid spans short-term (10/50) to long-term (150/200) combinations, covering both the commonly cited golden cross class and faster tactical entries.

Fast Period	Slow Periods Tested
10	50, 100, 200
12	50, 100
15	50, 100, 200
20	50, 100, 200
25	50, 100, 200
30	50, 100, 200
50	100, 150, 200
75	150, 200
100	150, 200
150	200

3. Results — Kill Filter Breakdown

25 variants tested. 0 survivors.

The failure splits almost evenly between two distinct modes — and the split matters for what to do next.

Kill Stage	Count	% of 25
Stage 1 — Net expectancy 0 or less	12	48.0%
Stage 2 — Effective sample size under 50	2	8.0%
Stage 3 — IS/OOS drift above 5pp	11	44.0%
Stage 4 — Binomial test	0	—
Stage 5 — Sharpe variance	0	—

Stage 1 (48%): Nearly half the variants never produce positive expectancy after 2-pip round-trip costs. These are primarily the slow-MA variants (100 and 200 bars) where entries arrive too late to capture meaningful residual momentum.

Stage 3 (44%): The remaining failing group clears the expectancy hurdle but collapses at the drift check. These signals have real gross edge — but the win rate learned in-sample does not transfer to out-of-sample windows consistently enough to trust.

Stage 2 (8%): Two variants — 25/200 and 75/150 — fire so infrequently that 6 years of out-of-sample data does not accumulate 50 effective trades. Statistical inference on these sample sizes is not meaningful.

4. All 25 Variants — Ranked by OOS Sharpe

Rank	Fast	Slow	OOS Sharpe	Net Exp (pips)	OOS Trades	Kill Stage
1	30	200	+0.862	+23.56	34	Drift
2	20	50	+0.837	+9.28	111	Drift
3	25	50	+0.770	+10.32	110	Drift
4	50	100	+0.709	+16.92	56	Drift
5	75	200	+0.571	+13.09	24	Drift
6	50	150	+0.563	+11.15	40	Drift
7	30	100	+0.390	+5.35	60	Drift
8	12	50	+0.309	+1.88	128	Drift
9	10	100	+0.271	-0.57	85	Net Exp
10	25	200	+0.158	+3.50	36	Eff Sample
11	30	50	+0.133	+2.53	112	Drift
12	150	200	+0.088	+0.52	32	Drift
13	75	150	+0.071	+2.14	39	Eff Sample
14	25	100	+0.027	-2.39	60	Net Exp
15	15	50	+0.015	+0.56	124	Drift
16	10	50	-0.008	-0.46	133	Net Exp
17	15	200	-0.052	-6.03	45	Net Exp
18	50	200	-0.108	-1.14	30	Net Exp
19	20	200	-0.122	-7.59	41	Net Exp
20	100	150	-0.136	-3.90	39	Net Exp
21	12	100	-0.139	-6.55	80	Net Exp
22	10	200	-0.186	-8.57	57	Net Exp
23	15	100	-0.350	-9.13	77	Net Exp
24	20	100	-0.471	-10.37	68	Net Exp
25	100	200	-0.707	-19.39	27	Net Exp

Blue = best performer. Green = sweet spot group (fast 25–75, slow 50). Red = worst performer.

5. Best Performer: SMA 30/200

A 30-period fast MA against a 200-period slow MA — a 6.7x separation ratio producing a low-frequency, trend-only signal.

Avg OOS Sharpe

+0.862

Avg Net Expectancy

+23.56 pips

Total OOS Trades

Kill Stage

IS/OOS Drift

At 34 total OOS trades across 6 years — roughly 5 to 6 trades per year — this is a highly selective signal. The aggregate numbers look encouraging. The per-window data exposes the reality.

Per-Window Breakdown

Win	OOS Period	OOS Sharpe	Net Exp	WR OOS	WR IS	Drift	Status
1	2015	+1.021	+37.92	60.0%	54.5%	+5.5pp	DRIFT
2	2017	+1.680	+61.86	80.0%	57.1%	+22.9pp	DRIFT
3	2019	+0.002	-1.97	42.9%	50.0%	-7.1pp	DRIFT
4	2021	+1.729	+31.12	80.0%	30.8%	+49.2pp	DRIFT
5	2023	+0.534	+11.04	40.0%	66.7%	-26.7pp	DRIFT
6	2025	+0.208	+1.37	71.4%	37.5%	+33.9pp	DRIFT
Aggregate		+0.862	+23.56	62.4%	49.4%	—	KILLED

Window 4 (2021), drift +49.2pp: IS win rate was 30.8%. OOS win rate was 80.0%. A 49.2pp swing. Five trades in the OOS window means a single win or loss moves the win rate by 20pp. The signal is not learning anything stable — it is fitting IS noise that happens to partially align with OOS performance.

Window 5 (2023), drift -26.7pp: IS win rate 66.7%, OOS win rate 40.0%. The reverse collapse in the following period. Statistical inference on 5 trades is coin-flip territory regardless of how many years of IS data preceded it.

The aggregate looks useful (+0.862 Sharpe, +23.56 pips). The per-window data reveals drift values ranging from -26.7pp to +49.2pp across six windows. A signal with that variance profile is not a deployable edge — it is variance that occasionally produces skill-looking outcomes.

Testing the Next Step

The 30/200 crossover has real gross edge — +23.56 pips average net expectancy — but unstable IS/OOS win rate transfer. The logical next test is conditioning the entry on a regime filter: only take the crossover signal when ADX is above 25, confirming a trend is present. Building that combination in MQL5 requires code most traders cannot write themselves. StratForge generates it from plain English in about 3 minutes.

Generate a custom SMA + ADX EA with StratForge

6. Worst Performer: SMA 100/200

The slowest combination in the grid. 27 OOS trades across 6 years. Loses money in five of six windows.

Avg OOS Sharpe

-0.707

Avg Net Expectancy

-19.39 pips

Total OOS Trades

Kill Stage

Net Expectancy

Win	OOS Period	OOS Sharpe	Net Exp	WR OOS	Drift	Status
1	2015	-1.354	-39.36	25.0%	-37.5pp	NET EXP
2	2017	-0.882	-22.60	25.0%	-8.3pp	NET EXP
3	2019	-0.516	-8.27	33.3%	-3.0pp	NET EXP
4	2021	-1.203	-23.78	20.0%	-30.0pp	NET EXP
5	2023	-0.555	-26.80	33.3%	+22.2pp	NET EXP
6	2025	+0.268	+4.47	25.0%	-41.7pp	NET EXP

Structural interpretation: A 100/200 crossover fires during confirmed macro trend changes — moves that have already travelled hundreds of pips before the signal triggers. By the time both averages agree on direction, the move is typically exhausted. The signal enters late into tired trends and exits on a fixed 8-bar timer, collecting whatever residual momentum remains. Against a 2-pip round-trip cost, late entries on tired trends produce losses by construction.

7. Parameter Analysis

Performance by Slow Period

Slow Period	Variants	Avg OOS Sharpe	Avg Net Exp (pips)	Drift Kills	Exp Kills
50	6	+0.342	+4.02	5	1
100	7	+0.063	-0.96	2	5
150	3	+0.166	+3.13	1	1
200	9	+0.056	-0.23	3	5

Slow=50 outperforms clearly. Average Sharpe +0.342, average net expectancy +4.02 pips. All failures in this group die at drift rather than expectancy — the signals have real gross edge that IS/OOS instability kills. This is a remediable problem.

Slow=100 and slow=200 are structurally weak. The majority of kills in these groups occur at Stage 1 (negative expectancy). These signals are not overfitting a marginal edge — they simply do not produce positive gross expectancy after costs.

Performance by Fast Period

Fast Period	Variants	Avg OOS Sharpe	Avg Net Exp (pips)
10	3	+0.026	-3.20
12	2	+0.085	-2.34
15	3	-0.129	-4.87
20	3	+0.081	-2.89
25	3	+0.318	+3.81
30	3	+0.462	+10.48
50	3	+0.388	+8.98
75	2	+0.321	+7.61
100	2	-0.421	-11.64
150	1	+0.088	+0.52

There is a clear fast-period sweet spot between 25 and 75. Fast periods below 20 produce negative or near-zero average net expectancy — too sensitive to intraday noise, generating frequent false crossovers with thin individual move capture. Fast periods of 25 to 75 produce the only consistently positive group. Fast=100 collapses to -0.421 average Sharpe — it converges too close to the slow period and fires too late into already-established moves.

8. Regime Analysis

Aggregating performance across all 25 variants by window reveals a consistent pattern — and one result that contradicts conventional SMA wisdom.

Window	OOS Period	Avg Sharpe	Avg Drift	Positive Sharpe
1	2015	+0.186	14.1pp	14/25
2	2017	+0.402	16.6pp	17/25
3	2019	+0.232	13.6pp	16/25
4	2021	-0.308	20.2pp	11/25
5	2023	-0.117	17.6pp	10/25
6	2025	+0.444	22.0pp	18/25

Windows 4 and 5 are hostile (2021 and 2023). Window 4 captures EUR/USD sustained multi-year decline — exactly where an SMA crossover should theoretically excel. Average Sharpe across all 25 variants: -0.308. The explanation is timing: 4H crossover entries arrive late in leg moves, after the trend is already visible to all participants. The profitable part of the move has passed by the time the crossover fires.

Drift does not improve in friendly windows. Even in Window 2 (average Sharpe +0.402), average drift is 16.6pp. The signals overfit their IS win rates regardless of whether the OOS period is favourable or hostile. Regime-friendly conditions improve the absolute Sharpe but do not reduce the IS/OOS instability that the kill filter correctly identifies.

9. The Structural Problem with SMA Crossovers on 4H FX

1. The 4H Bar Contains Too Much Noise for Trend Confirmation

SMA crossovers were developed on daily data, where each bar represents a full trading session and noise is partially averaged out. On 4H, each bar is a single session segment. The signal-to-noise ratio is materially lower. A fast MA crossing a slow MA on 4H is frequently reacting to session-specific flow rather than genuine directional change. The signal fires, the session-specific move reverses, and the 8-bar exit captures a retracement rather than a continuation.

2. Transaction Costs Consume the Edge at High Signal Frequency

The variants with the most frequent signals — fast=10, 12, 15 against slow=50, producing 120 to 133 OOS trades — are exactly the variants with the worst average net expectancy. At 2-pip round-trip, a signal that averages 3 to 5 pips gross per trade is left with 1 to 3 pips net. Any reduction in win rate or average move size in an OOS window tips net expectancy negative. The cost-to-edge ratio for high-frequency SMA crossovers on 4H is structurally unfavourable.

3. The In-Sample Window Blends Multiple Regimes That Do Not Persist

A 2-year IS window on EUR/USD 4H contains trending periods, ranging periods, and regime transitions. The SMA parameters that best fit this blend are not the parameters that best fit any specific future regime. When the OOS year presents a single dominant regime, the IS-optimised parameters are typically miscalibrated for it. This is what the large IS/OOS drift values — mean 17.3pp across all windows — are measuring: systematic IS-to-OOS parameter misalignment driven by regime composition shifts between windows.

The Logical Next Step

The pattern from Part 1 points toward an obvious hypothesis: if SMA crossovers have real gross edge in the 25 to 75 fast period range but fail on IS/OOS drift, can a second indicator stabilise the win rate transfer? Testing that manually means choosing indicator pairs, parameterising both, and running walk-forward validation on every combination. That is weeks of work per hypothesis. StratForge generates the MQL5 EA code for combination strategies from plain English — describe the logic, receive the code, run the backtest in MetaTrader 5.

Generate a custom combination EA with StratForge

10. Cross-Series Comparison

This table accumulates with each instalment. All entries use identical methodology.

Part	Indicator	Variants	Mean Drift	Drift Kill %	Survivors	Dominant Kill
1	SMA Crossover	25	17.3pp	44%	0	Net expectancy (48%)
2	MACD	75	See Part 2	See Part 2	0	See Part 2

11. Frequently Asked Questions

What is the best SMA crossover setting for MT5?v

Our 13-year walk-forward backtest found the 30/200 SMA crossover to be the best-performing of 25 configurations, with an average OOS Sharpe of +0.862 and net expectancy of +23.56 pips. However, it failed the IS/OOS drift check in every single validation window. The fast period sweet spot across the parameter grid is 25 to 75 bars against a 50-period slow MA, which produced the only consistently positive expectancy group.

Why do SMA crossover strategies fail in live trading?v

Two distinct failure modes. First, slow moving averages (100 and 200 bars) lag so severely that entries arrive after most of the directional move is spent. Second, faster configurations that do produce positive gross expectancy fail the IS/OOS drift check: win rates learned in a 2-year in-sample period do not transfer to the following year because the specific regime mix shifts between windows. Mean drift of 17.3pp across 150 validation windows is not an edge — it is variance.

Does the golden cross (50/200) SMA work on forex?v

No, not on EUR/USD 4H under realistic conditions. The 50/200 crossover falls into the slow-MA group that fails at Stage 1 — negative net expectancy after 2-pip round-trip costs. The 100/200 crossover, the slowest tested, produced the worst result in the study: average Sharpe minus 0.707, average net expectancy minus 19.39 pips.

What SMA period combination shows the most promise?v

Fast periods of 25 to 75 bars against a 50-period slow MA produced the only consistently positive expectancy group — average OOS Sharpe +0.318 to +0.462 and average net expectancy +3.81 to +10.48 pips. The logical next step is combining with a regime filter such as ADX above 25.

Is SMA or MACD better for trend following on MT5?v

SMA crossovers showed a lower drift kill rate (44%) than MACD (90.7%). Nearly half of SMA failures occur at the net expectancy stage — the signal simply loses money outright — whereas MACD failures are dominated by regime overfitting. Full comparison data is in Part 2 of this series.

12. Next in the Series

This is Part 1 of the MT5 Indicator Backtest Series. Every study uses the same walk-forward methodology, the same dataset, and the same 5-stage kill filter — making results directly comparable across indicator families.

Part 1 — You are here

SMA Crossover Analysis

25 variants · 13 years · 0 survivors · Split failure: 48% expectancy, 44% drift

Part 2 — Read next

MACD Configuration Analysis

75 variants · Same methodology · Drift dominates at 90.7% kill rate

Build on This Research

The fast 25 to 75 against slow 50 group has genuine gross edge that fails on IS/OOS drift. The natural next step — combining with a regime filter or second confirmation signal — requires multi-indicator MQL5 code most traders cannot write themselves. StratForge generates it from a conversational strategy interview. No coding required. Your own Anthropic API key means you pay Anthropic directly — typically a few cents per generation.

Generate a Custom SMA EA