Gold on M1 moves enough every day to look like free money. Then you run a scalper on it for a month and find out where the money went: spread, commission and a few dollars of slippage on every single trade. Most M1 gold EAs aren't wrong about direction. They just trade too often for whatever small edge they have to survive the costs.

Before you read further: I built Oceanus and I sell it, on the MQL5 Market and on GogoJungle. So this isn't an independent review. The numbers below come from my own simulation on real Exness ticks and from the MT5 Strategy Tester, and they're backtests unless I say otherwise. There's also a live signal on a real account that you can check yourself.

Updated October 23, 2026 for version 1.33. That version enters with pending stop orders and uses retrained models, so I re-ran every number, chart and table in this post on the new logic. The old figures are gone, not mixed in.

The product pages are short by design. This post covers what didn't fit there: what the EA actually does all day, how to read the backtest, and which account and deposit make sense. That last part matters more than I expected.

What a selective gold scalper EA actually does

Most of the time, nothing.

Oceanus has three layers, and only one of them is about entering trades:

  1. Timing. Two machine-learning models run inside the EA and look at the gold tick stream once a minute. Their only job is to say whether the next stretch of time is worth trading. Most of the time they say no.
  2. Direction. When the models allow it, the EA places a Buy Stop above the market and a Sell Stop below it. Whichever side price breaks first gets filled, and the other order is cancelled.
  3. Management. Stop loss from the first tick. A quick move to breakeven, set $0.30 in profit per 0.01 lot so commission is covered. A trailing stop that only tightens. And a time limit, so a trade that hasn't done its job gets closed instead of left open for days.

I'm not going to publish the exact rules or the model inputs, for the obvious reason. But the split matters, because it explains how the EA behaves. The models never decide buy or sell. I tested that: models asked to predict direction from price features scored no better than a coin flip (AUC 0.47 to 0.52). Models asked whether gold is about to move a lot did much better. On 2026 data, which they never saw during training (January to October 22), they were right 75.7% of the time, against 65.8% for always guessing the most common answer. They were sharpest in the wild months at the start of the year and weaker as gold calmed down. By October they were barely better than that constant guess, which is one more reason quiet gold means few trades.

So the ML picks the moment and the rules handle the rest.

In practice that means a median of 3 trades on days when it trades and up to 34 on a wild day. How many days it trades depends entirely on gold: over the whole test more than half of all weekdays had no trade, but in busy 2026 only about one weekday in fifteen went without one. Median time in a trade is 8 minutes, and 94% of trades are closed within an hour.

How I got here (the short version)

Oceanus didn't start as a breakout system. It started as an attempt to make a grid EA safer by letting the models decide when the grid could open. I simulated that across 4.7 years of ticks and the grid made roughly zero before costs, whatever the models said. After costs it lost. So I dropped the grid and kept the models.

The part that took the longest wasn't the strategy, honestly. It was getting the models inside the .ex5 to give exactly the same answers as my Python originals, minute by minute. They now match to the seventh decimal place. The trades themselves don't match quite as tightly. I ran the finished EA in the MT5 Strategy Tester on 100% real ticks from February 13 to March 20, 2026, with 750 ms of simulated latency: 201 trades, +$584 at 0.01 lot, profit factor 2.15. My research simulation took 225 trades and made +$533 over the same weeks. Same story, not the same trade list. And those five weeks were among the best of the whole history, so don't read them as typical.

One decision that surprised me: the time limit. About 7% of trades end because time ran out, and as a group those trades lose money (-$1,275 at 0.01 lot over the full test). The obvious fix is to drop the limit and give them more room. I tested that on October 2, on the earlier build. It was worse. Trades that are still red after the window tend to go on and hit their original stop. Cutting them early is part of the edge.

Backtest results, read properly

Conditions for everything in this section: XAUUSD, Exness Raw Spread tick data, real spread on every tick, commission $0.04 per 0.01-lot round trip, fixed 0.01 lot, current trading logic (v1.33). January 2022 to September 16, 2026. Pending orders are filled on the tick that reaches them.

Year What gold did Trades Result at 0.01 lot
2022 choppy, rate hikes 295 -$41
2023 quiet range 75 +$47
2024 rally starts 266 +$77
2025 strong trend 510 +$136
2026 (to Sep 16) big swings 1,312 +$1,475
Total 2,458 +$1,694, profit factor 1.25

Equity curve, Oceanus on XAUUSD M1, Jan 2024 to Sep 16 2026, fixed 0.01 lot from $2,000, Exness Raw ticks

From January 2024 on (the period in the charts): 2,088 trades, +$1,688, profit factor 1.27, max drawdown $221, which is 10.2% measured from the equity peak on a $2,000 start.

Look at the yearly table again, because it tells you what kind of EA this is. It earns when gold moves. From 2022 to 2025 it barely did anything useful: a small loss in 2022, small gains after that. Then gold had its busiest stretch in years, and nearly all the profit came from October 2025 to March 2026. If gold goes quiet for a year, expect a flat year.

Profit by quarter at 0.01 lot, Oceanus on XAUUSD M1, 2024-2026 backtest

Monthly profit heatmap in USD at 0.01 lot, Oceanus on XAUUSD M1, 2024-2026 backtest

Over the whole test, 30 months closed green and 25 red. Worst month: November 2025, -$59. Best: February 2026, +$523.

Why a 79% win rate and only PF 1.25? Because most wins are small. 78% of trades close at a stop that's already been moved to breakeven or trailed, averaging +$4.18. Only 8 trades out of 2,458 ever reached take profit. The losers that hit the original stop are 14% of trades, and they average -$15.42. In 2024-2026 the average win was $4.77 and the average loss $14.50. So the win rate is high because the breakeven locks in a little on most trades, not because the EA is clever at picking winners.

Trade analysis for Oceanus on XAUUSD M1, 2024-2026 backtest: P/L distribution, win rate, key statistics

The 10 worst trades

I publish these for every EA I test. All at 0.01 lot, times in UTC.

# Entry Exit Side Entry price Exit price P/L Exit reason
1 2026-01-30 15:05 15:18 Sell 4981.44 5058.63 -$77.23 original SL
2 2026-03-09 01:01 02:10 Sell 5018.09 5091.34 -$73.29 original SL
3 2026-02-05 08:13 09:04 Buy 4941.19 4883.23 -$58.00 original SL
4 2026-02-05 16:54 18:16 Buy 4902.92 4851.70 -$51.26 original SL
5 2026-06-24 13:13 13:42 Sell 3967.06 4017.13 -$50.12 original SL
6 2026-02-04 14:39 14:54 Sell 4971.54 5017.84 -$46.35 original SL
7 2026-03-27 15:23 17:02 Buy 4551.73 4507.75 -$44.02 time limit
8 2026-03-20 11:01 12:01 Sell 4638.15 4681.86 -$43.75 original SL
9 2026-03-24 13:34 14:10 Sell 4355.67 4399.04 -$43.41 original SL
10 2026-02-05 01:25 01:37 Sell 4940.00 4981.95 -$41.98 original SL

This is the weak spot, and I'd rather you see it here than find it on your account. One of these trades wipes out 10 to 20 small winners. Nine of the ten happened between late January and March 2026, when gold was swinging $100+ in an hour. The stop is placed by market structure, not a fixed number of pips, so in a violent market it sits far away. The EA survived that stretch and finished 2026 well up, but those weeks were rough.

Costs. Add up to $0.10 extra cost to every single trade (wider spread, higher commission or slippage) and the profit factor stays above 1.20.

Cost sensitivity for Oceanus: extra cost per trade vs profit factor, XAUUSD M1 backtest 2024-2026

If you want to check any of this, Oceanus is on the MQL5 Market with a free demo for the Strategy Tester. Backtesting instructions are further down.

Which account type should you run it on?

A raw-spread account. The account choice moves the result more than any input setting does.

Same EA, same rules, same gold, run on tick data from different account types. First, Exness Raw against Exness Standard Cent over the full period:

Account Trades 2022-2024 2025-2026 Total Profit factor Max DD
Exness Raw 2,419 +$83 +$1,600 +$1,682 1.25 $221
Exness Standard Cent 2,405 -$61 +$1,414 +$1,353 1.20 $249

v1.33, Jan 2022 to Sep 16 2026, results expressed for a standard 0.01 lot (on a cent account the same figures are in US cents at 0.01 cent lot). March 30-31, 2026 is excluded for both accounts because the Exness tick archive for the cent symbol is missing those two days.

The gap in gold spread makes a cent account cost roughly $0.10 to $0.20 more than Raw on every trade, even after you count Raw's commission. Gold spread on the cent account ran $0.16 to $0.36 in the months I checked, against $0.04 to $0.14 on Raw. That's about $0.14 less profit per trade, and with an edge this thin it's the difference between 2022-2024 at -$61 and the same three years slightly positive on Raw.

Then a wider comparison over a shorter window, including a second broker:

Account Total, Apr 2025 to Sep 16 2026 Profit factor Max DD
Exness Raw +$1,604 1.28 $221
Exness Zero +$1,535 1.27 $248
Pepperstone Razor +$387 1.06 $436
Pepperstone Standard +$172 to +$230 1.03 $495-516

v1.33, 0.01 lot. Commission per 0.01-lot trade: Raw $0.04, Zero $0.11, Razor $0.08, Standard none. Pepperstone Standard was simulated as Razor prices plus a 0.10-0.13 markup.

Pepperstone surprised me. On the earlier build, which entered with market orders, it lost money over this window. Pending stop orders brought it back above zero, but it still makes about a quarter of what Exness makes, with twice the drawdown. Part of that is cost (wider spread, higher commission). The bigger part, measured on the earlier build, is how its price moves at the exact moment of a breakout: Pepperstone's quotes jump further past the level than Exness quotes do, worst around US news at 12:00-15:00 UTC.

My take, account by account:

  • Exness Raw or Zero: what I use and what all the main numbers are based on. Raw has the more predictable cost.
  • Another broker's raw/ECN account: possibly fine, but backtest on that broker's own real ticks first. Low commission alone isn't enough, as the Pepperstone line shows.
  • Exness Standard Cent: it works, more weakly. Good for running it live on small money while you watch it. Not where I'd expect the full result.
  • Standard accounts with wide gold spreads: I wouldn't.
  • Prop firm accounts: stop loss on every trade and one position at a time suit most rules. There's no built-in daily loss limit, though, and a single bad trade can reach $77 at 0.01 lot, so size against your firm's daily limit.

How much capital, and what profit is realistic?

At 0.01 lot on a raw account, I'd start with $2,000. That put the worst drawdown in the test at 10.2%.

Setup Deposit per 0.01 lot What the backtest says
Conservative $2,000 max DD 10.2% (2024-2026)
Moderate $1,000 same $221 drawdown, now roughly 20% of the account
Compounding add 0.01 lot per $1,000 of balance $10,000 to $46,050 over 4.7 years, max DD 20.4% (closed trades)
Cent account $50 worst DD in the test was 249 cents ($2.49), so the deposit isn't the issue, the thinner edge is

On profit, think in dollars per 0.01 lot per year rather than a monthly percentage, because the answer depends on gold. In a quiet year the backtest made somewhere between -$41 and +$77 per 0.01 lot. In a busy year, from a hundred to well over a thousand. If someone promises you a fixed monthly return from a system like this, they're guessing.

I deliberately didn't build in "% risk per trade" sizing. I tested it from $2,000: at 0.5% risk per trade it ended almost exactly where the fixed lot did ($3,725 against $3,694) with the same 10% drawdown. At 1% it ended at $3,970 with the drawdown doubled to 21%, and at 2% at $4,282 with 38%. A little more profit for a lot more pain. Small stops make small risk-based lots, and commission is a fixed cost, so the tight-stop trades become expensive. A fixed lot that you raise yourself works better here.

Getting the most out of it

  • Use a VPS close to your broker. Entries are pending orders that the broker fills on its side, but the EA still moves those orders every minute and trails the stop many times a day, all from your terminal. My VPS in Vietnam pings the Exness server at 63 ms. The first one I used, in Germany, was at 190 ms.
  • Set InpServerUtcH correctly. It's your broker's server time minus UTC (0 for Exness). It's the one setting people skip, and it shifts everything the models see.
  • Start small. Two to four weeks on demo or at 0.01 lot before you size up.
  • Read the slippage log. The EA writes every fill against its expected price to Common\Files\Oceanus. If your fills are consistently worse than about $0.10, your broker or your VPS is eating the edge.
  • Don't switch it off in a drawdown. The profit comes in a few busy months. Pausing after a bad week is a good way to miss the month that pays for the year.
  • Raise the lot after you've sat through a real drawdown, not after a good week.

For backtesting: set InpModels to Exam, use "Every tick based on real ticks", and start the test 16 days before the period you care about so the EA can warm up. The Exam models were trained only up to the end of 2025, so a 2026 test is genuinely out of sample. That's the one I'd look at first.

Questions you're probably asking

"Almost all the profit is from 2025-2026. Isn't that overfitting?"
Fair question. Two things argue against it. The 2026 results come from models that never saw 2026 data. And the pattern has a plain cause: the EA trades movement, and 2025-2026 had the most movement. What I can't promise is that gold stays this busy. If it goes quiet, expect results closer to 2023.

"How is the live account doing?"
The live signal has run on a real Exness Raw account at a fixed 0.01 lot since October 1, 2026, on the earlier build until October 23 and on the current pending-order logic since then. That's weeks, not months, so it doesn't prove anything yet. I'm using it to measure real slippage against the backtest. I'll update this section as the record builds.

"Why didn't it trade today?"
The models didn't like the market. The panel on the chart tells you whether it's waiting or armed.

"Can I run it on another symbol or timeframe?"
No. The models were trained on gold M1 and nothing else.

If you're still reading

You probably care more about how the numbers were made than about the numbers themselves, and that's the reader I write for. I review other people's strategies on this blog the same way. The TradingView SuperTrend strategy backtest and the follow-up on optimizing SuperTrend for gold are good examples: one setup looked great right up until I shifted the candle open by an hour.

Oceanus is available on the MQL5 Market and on GogoJungle. Updates are free, and every change to the trading logic gets re-run on the full tick history before it ships.

Last updated October 23, 2026 (version 1.33).


Risk warning: trading gold and CFDs carries a high level of risk. All figures above are backtests on historical data unless stated otherwise. Past performance does not guarantee future results. Test on a demo account first and only trade money you can afford to lose.