Gold GodTier Dual Neural
- Experts
- Version: 88.69
- Updated: 13 September 2026
- Activations: 10
GOLD GODTIER - Dual ONNX AI trend system for XAUUSD M15
WHAT IT ACTUALLY IS
- Two ONNX models (LSTM + XGBoost) embedded in the file. No DLLs, no internet calls, nothing to download.
- ONE position at a time. Fixed ATR-based Stop Loss and Take Profit on every single trade.
- NO martingale. NO grid. NO averaging. NO recovery mode. Ever.
- Friday-close protection (flat before weekend gaps), hard equity-stop guard, consecutive-loss circuit breaker, rollover-spread block.
- 9 one-click presets: attach, pick a preset, done.
WHAT I WILL TELL YOU THAT OTHER SELLERS WON'T
This system wins about 63% of trades with the average win slightly SMALLER than the average loss (payoff 0.65). It earns by frequency, not by home runs. About those backtests, so you can weigh them properly: 2014-2024 was used to develop and tune this system, so those years are IN-SAMPLE and prove little on their own. 2025 and the first half of 2026 were never used to tune parameters. But the two ONNX models inside the file are a separate question, and I had this wrong on this page until I measured it. Judged by their own year-by-year accuracy, the training window of the models runs from January 2017 to December 2025, so 2025 is in-sample for the models even though I never tuned on it. That leaves 2015, 2016 and 2026 as the only years the models had genuinely never seen. 2015 and 2016 both lost money. 2026 was profitable, but the models showed no measurable predictive power that year, so it is more likely gold rising than model skill. The screenshots mark the training window so you can see this yourself. The honest summary is that the long backtest is mostly in-sample, and the real test is the live record, which is only starting now. Everything before 2022 is bar-model only, because real tick data does not exist that far back on my broker. From the year-by-year backtests:
- About 30% of any 8-week window is a LOSING window. This is normal for this system. If you cannot sit through 8 red weeks, please do not buy this.
- Longest losing streak you should expect: 6-7 trades.
- Quiet periods of 3-10 trading days without a single trade are normal.
- The best 100 trades out of about 1,900 carry roughly a third of the gross profit. Take those away and twelve years comes out flat - profit factor 1.00. The edge is real, but it is concentrated: most of the money comes from a minority of the trades. Plan for that.
THIS SYSTEM IS DIRECTIONALLY ASYMMETRIC - READ THIS BEFORE ANY NUMBER
In a sustained uptrend this EA trades LONG ONLY. I measured that instead of assuming it.
Across 15,966 decision bars in 2025 on gold, the model's buy confidence stayed inside a narrow band: median 0.497, 1st percentile 0.447, lowest reading of the entire year 0.335. A short entry requires that confidence to fall below 0.41. It did so on 0.16% of bars - 26 out of 15,966 - and the local trend filter blocked every single one. The result was ZERO short trades in the whole of 2025.
Gold rose through 2025, so long-only was the profitable side that year. That makes 2025 a trend capture, not evidence of a two-sided edge, and you should weigh it that way. In the first half of 2026, when gold was less one-directional, the same preset took 22 short trades out of 277.
If you want a robot that shorts gold regularly, this is not it. I would rather you know that before you buy than find out in month two.
NEW IN 88.55 - AN OPTIONAL CORRECTION TO THE SELL SIDE (DEFAULT OFF)
Having told you the above, I went looking for the cause. It is not the models. Taking those same 15,966 decision bars, sorting them by the model's buy confidence and measuring what price actually did over the following hour, the difference between the extremes is real and tradeable: the lowest-confidence tenth of bars fell on average, the highest-confidence tenth rose, a spread of about 2.0 USD per ounce against a 0.16 spread cost, t = 4.5. The information needed for short entries was there all along. The sell threshold was simply set deeper than the model's output range ever reaches, so roughly 98% of it was discarded.
88.55 adds one input, InpSellThreshOffset, which lowers the sell threshold by the amount you give it. It defaults to 0.0, which reproduces 88.54 exactly - if you change nothing, nothing changes, and I verified that with a parity test that matched to every digit.
At 0.065, the value that puts both thresholds at the same percentile of the model's output distribution, the backtests move as follows. 2026 first half on real ticks: short trades 7% to 56%, return 7.5% to 37.6%, maximum equity drawdown 11.3% down to 7.5%. 2024 on the bar model: short trades 0.7% to 27%, return 22.4% to 35.1%, drawdown 5.7% down to 4.8%.
Apply the same scepticism I asked of you above. Weigh that carefully. Only the 2026 first-half run is genuinely out-of-sample - about six months. The 2024 run is inside the models training window, so it is a consistency check, not independent evidence. And there is almost no live history behind any of it. I am running it on my own money from 11 September 2026 and I will report what happens either way. Until there is a live record worth showing, the default stays off and I am not asking you to turn it on.
TRACK RECORD - WHERE IT STANDS RIGHT NOW
The public signal this page used to link to is no longer available, so I have removed the link rather than leave a dead one pointing at proof that is not there.
What I can show you today is backtests, and I have told you above exactly which part of them is out-of-sample and which is not. Please weigh this product on that basis, and on the demo you run yourself on your own broker. When a new public signal has real trades behind it, it will be linked here.
GROWTH+ (Preset = 8) - NEW IN 88.50
For accounts of 3,000 USD and above. GROWTH+ is not a new strategy: it is the GROWTH preset with a higher risk per trade (2.50% instead of 1.70%). Signal, filters and every protection stay identical. Larger positions mean larger profits and larger drawdowns in the same proportion, and the hard equity stop stays at 20% - which is why the account size matters. On anything under 3,000 USD use GROWTH (Preset = 2); the EA will warn you in the Experts tab if you do not.
NEW IN 88.52 - THE EXIT IS NOW OPEN AND TESTED
The two partial-close stages used to be fixed inside the code. They are now real inputs that work in every preset: InpPartialR1 (stage 1 trigger, in R multiples), InpPartialPct1 (stage 1 close percent), InpPartialR2, InpPartialPct2.
The defaults - 1.0R close 30 percent, then 2.0R close 50 percent of the remainder - are exactly what the EA has always done, and they are what I recommend. I swept 12 variants over 12 continuous years, then re-tested the two best ones over 12 separate single years: none beat the defaults once the compounding effect was removed. Change these only if you have run your own year-by-year test on your own broker data.
The old InpPartialATR input never affected anything. It is now labelled NOT USED and kept only so existing .set files keep loading.
BEFORE YOU BUY - TEST THE DEMO YOURSELF
Run the free demo in your own Strategy Tester, on YOUR broker's data, with YOUR spread, using "Every tick based on real ticks". My numbers on my broker are irrelevant to you; yours are what matter.
CHECK YOUR BROKER'S GOLD SPREAD BEFORE YOU RUN IT
This is the single most common reason the EA appears to "do nothing".
The EA has a hard maximum-spread filter and it is deliberate: entries taken while the gold spread is abnormally wide were measured to be unprofitable, so the EA stays out instead of paying that cost.
Maximum spread per preset (gold, 2-digit quotes, value in points):
SAFE (the default) - 45 points
GROWTH - 50 points
GROWTH+ - 50 points
BEAST - 40 points
If your broker's live gold spread sits above the limit of the preset you selected, the EA will open NO trades at all and the on-chart panel will display SPREAD HIGH (xx.x > yy.y). That is not a malfunction - it is the filter doing its job. For reference, the average gold spread on my own broker is currently around 56 points, which blocks the default SAFE preset completely.
Recommended: an account whose average gold spread is 35 points (0.35 USD) or lower. Many brokers offer a lower-spread account type on the same login (often called Raw, Zero or Ultra Low). Gold spreads also widen at rollover and around major news. Open Market Watch and look at your own live gold spread before you run it - the dashboard always shows the current spread next to the active limit, so you can tell at a glance whether the EA is able to trade.
WHAT I MEASURED IN SEPTEMBER 2026 - THE COST NUMBER THAT DECIDES EVERYTHING
I stopped guessing about spread and measured it. These are the numbers.
First, how much the cost actually matters. I rebuilt the EA's entry logic outside the terminal - the same two models, the same 0.59 threshold, the same trend bias and the same local EMA filter - and checked that it reproduced the real thing: about 275 signals a year, roughly 204 trades after the one-position rule, most of them long. It matched the live backtests. Then I re-ran exactly those trades at different gold spreads:
- 31 points: expectancy +0.115 R per trade, t = 6.5. A clear, measurable edge.
- 40 points: +0.075 R, t = 4.2.
- 50 points: +0.037 R, t = 2.1.
- 56 points: +0.017 R, t = 0.9, p = 0.35. No longer distinguishable from zero.
The edge does not disappear at one cliff - it bleeds out as the spread widens. At a standard gold spread, what is left is inside the noise. At a low-spread account it is not. That is the whole reason spread is the first thing I tell you to check.
Then I measured real spreads by reading tick history straight out of the terminal, over 7 million ticks per account, all sampled at the same moment on the same broker:
- Standard-type gold account: 56 points average over 30 days. Only 0.15% of ticks were at 50 points or below.
- Low-spread account type (Raw / Zero / Ultra Low): 30 points average. 99.5% of ticks at 50 points or below.
Same EA, same settings, same market, same minute. On one account the costs consume the edge; on the other they do not. This is also why the EA on a standard gold account often shows SPREAD HIGH and never trades. It is refusing to pay a cost it cannot recover, which is exactly what it should do.
A MICRO GOLD CONTRACT LOWERS THE DEPOSIT YOU NEED
The $3,000 minimum exists because standard gold is a 100 oz contract with a 0.01 minimum lot, so the smallest position available is 1 oz - too large for the risk engine to size correctly on a small account.
Several brokers also list a micro gold symbol: 1 oz per lot with a 0.10 minimum lot, so the smallest position is 0.1 oz - ten times finer. On that contract the risk engine reaches its intended size accurately at far smaller deposits, and the account stops forcing the EA to risk more than it is asking for. Spread and swap on those micro symbols are usually identical to the low-spread standard symbol - I measured 30.69 against 30.70 points on the same broker.
The strategy's edge does not change with account size. What changes is whether the risk engine is able to express it. If your deposit is below $3,000, look for a low-spread MICRO gold symbol rather than simply adding more money.
CENT ACCOUNT SUPPORT - NEW IN 88.54
The same reasoning as the micro contract above, taken further. On a cent account one lot is 1 oz and the minimum lot is 0.01, so the smallest position is 0.01 oz - one hundred times finer than standard gold. The risk engine reaches its intended size accurately on a deposit of a few hundred dollars instead of $3,000.
From 88.54 the EA reads contract size, digits and account currency at startup and recognises these accounts, instead of reporting a spec problem as earlier builds did. Attach it and check the Experts tab: it must print ENV VERDICT: OK.
Verified on Exness Standard Cent. Other cent brokers should work but I have not tested them, and I have no live-money track record on a cent account yet. Backtests on that symbol only cover 2024 to mid-2026, and the directional warning above applies there as well.
REQUIREMENTS
XAUUSD (Gold) - M15 - one chart - hedging or netting account. Minimum deposit $3,000. Leverage 1:100 or higher. VPS recommended. Gold spread of 35 points or lower strongly recommended - see the spread section above.
WHY THE MINIMUM WENT UP FROM $2,000 TO $3,000 IN 88.53
Gold trading above $4,000 an ounce makes the 0.01 minimum lot large relative to a small account, so the risk engine cannot size down far enough. I re-tested 2026 year by year at several deposits. At $2,000 both GROWTH and GROWTH+ reach an equity drawdown above 20% - that is the hard equity stop, and hitting it shuts the EA down permanently. At $3,000 the same period comes in at 14.8% (GROWTH) and 13.2% (GROWTH+). Eleven of the twelve tested years stay near 13% at any deposit; 2026 is the year that breaks a small account, because gold now trades at roughly four times the price it had when the 0.01 lot floor was set.
I would rather lose the sale than have you hit the hard stop in month three. If you have less than $3,000, please do not buy this yet.
WHAT THIS IS NOT
Not a get-rich machine. Not "no-loss AI". Expected long-run profile, measured on real ticks with the exact preset that ships (XAUUSD cent: 2024, 2025, 2026H1): profit factor 1.19-1.35 with maximum equity drawdown 5.5-11.2%. The higher figure I used to quote here came from twelve years that are mostly 1-minute bar-model data, and on the same configuration real ticks cut the result by about a third. Real ticks are the honest number and they are what I quote from here on. Some years were barely breakeven. If a seller promises you more than this, ask them for their year-by-year results.
The price increases after each profitable quarter on the public signal. Early buyers get the lowest price this product will ever have.
NETTING ACCOUNTS
Partial profit taking uses an API that exists only on hedging accounts, so it is skipped on netting accounts; break-even and the trailing stop still run. On a netting account this EA needs exclusive control of the trading symbol: do not trade the same symbol by hand and do not run another EA on it. Results tested on hedging accounts will not be reproduced exactly on a netting account.
