Polaris Edge

Polaris Edge is a multi-model machine-learning EA built specifically for XAUUSD on the M5 timeframe. It runs six ONNX classifiers in parallel, one per (strategy, side) combination, on four independent magic numbers. Up to four positions can be open at once on the same symbol, each managed independently with its own stop, take profit and trailing logic.

The point of having four magic slots instead of one is that the sub-systems are uncorrelated by design. They can be all long, all short, or mixed, depending on what each model sees on the last closed M5 bar. Two of the six slots are off by default (their thresholds are pinned at 0.95) because out-of-sample testing on the slot's data was weak and I'd rather ship it conservative than have you find out the hard way.

How entries work

At every M5 bar close, the EA builds a 26-feature snapshot from M5, M15 and H1: RSI(14), RSI(7), Bollinger z-score, ATR(14), relative position over 20 and 96 bars, 1-bar and 4-bar returns, plus current hour and day of week. That vector goes through all six models. Each model is a binary classifier outputting a probability between 0 and 1; if it crosses the per-slot threshold and the slot isn't already holding a position, a trade is opened.

Stops are wide on purpose (260 pips default, which on XAUUSD is $26 per 0.01 lot). The protective layer is the trailing stop, not the SL itself. As soon as price moves in your favor by enough to engage the trailing logic, the SL ratchets forward and most trades end up closing well inside the original SL band. You can run the trailing in two modes: fixed pips (default — 12 pips to engage, 3 pips step) or ATR-adaptive (start at 0.5 × ATR(M5,14), trail at 0.3 × ATR). Fixed-pips is the shipped default; the ATR mode is provided for traders who want the trailing distance to scale with realized M5 volatility.

P5 - a fifth, independent strategy (new in v2.1)

Alongside the four ML slots, Polaris Edge now runs a fifth strategy, P5, on its own magic number. It doesn't share magic numbers, thresholds or position limits with P1-P4, and it can hold a position at the same time as any of them.

P5 has its own risk preset - Low / Medium / High, the same balance-scaling model used for the main position sizing below - and ships live on the High preset by default.

P5 adds its own hourly order traffic on top of the M5 tick-driven ML pipeline, so execution speed to the trade server matters even more than it already did. See the broker note above.

Position sizing

Lot is recomputed on every new trade from AccountBalance, divided by a factor that depends on the risk preset:

• Low — balance / 200,000 (effective leverage roughly 1:2.4)
• Medium — balance / 100,000 (roughly 1:4.7, this is the default)
• High — balance / 50,000 (roughly 1:9.4)

There's also a Fixed mode if you want to set the lot by hand for backtesting or for staying on 0.01 deliberately. A hard cap (default 5.0 lots) protects against accidentally over-sizing when balance gets large. When the computed lot would fall below the broker's minimum (typically 0.01), the EA simply floors it to that minimum — small accounts trade at minimum size until balance grows enough for the preset to lift sizing above the floor.

The three presets weren't pulled out of a hat. I ran a Monte Carlo simulation, 2,000 iterations over a 12-month forward horizon, including a stress scenario where winning trades are haircut 50% to simulate model drift or an unfavorable regime. Low survives both with zero margin-call probability. Medium survives the current regime cleanly and has roughly a 12% margin-call rate in stress on tight balances. High doubles the throughput but in stress its margin-call probability is around 45%, so use it only with a big enough cushion.

Practical rule, if you want a single number: balance equal to or greater than $100,000 multiplied by your lot size. For 0.10 lot that's $10,000, for 0.50 lot that's $50,000. That keeps the effective leverage tight enough to absorb a bad month.

• $1000 : lot 0.01
• $2000 : lot 0.02
• $5000 : lot 0.05

Backtest, the honest version

The historical backtest I'm shipping is XAUUSD.r M5, real-tick mode, from November 2024 to May 2026 — 18 months. Numbers: 2,160 trades, win rate 96.7%, profit factor 2.09, max drawdown around 5% of initial balance, roughly $2,005 net profit on a $5,000 starting account.

Out-of-sample test on Jan–May 2026 (after the training cutoff): profit factor 2.55. Better than the training-window PF, which is unusual but real on this dataset.

The edge has also been validated statistically: I ran Monte Carlo null tests against random-entry and random-direction baselines (100 iterations each). On all three metrics — profit, profit factor and win rate — the result is significant at p < 0.01, confirming that both the entry timing and the direction selection contribute real predictive value rather than the EA riding stop-loss geometry by luck.

The on-chart panel

There's a live dashboard in the top-left corner showing the current risk preset, balance, equity with percentage change, computed lot for the next entry, margin used vs free, open positions with color-coded P/L per ticket, session stats, today's drawdown, max DD run, spread vs the configured maximum, countdown to next M5 close, and the live ML probability for each of the six slots with a flag when one crosses its threshold. It updates once per second. You can turn it off from the inputs if you prefer a clean chart.

What you need

MetaTrader 5 build 3815 or higher (ONNX runtime). Hedging account. XAUUSD or your broker's gold symbol (XAUUSD.r, XAUUSD.m, etc., works fine as long as spread is reasonable). Leverage 1:30 or higher. A VPS is strongly recommended — the EA reacts at every M5 close and missing ticks during a fast move will delay entries.

The EA always trades from 0.01 lot regardless of preset. The preset only kicks in to scale lot up once the balance is large enough. I'd recommend starting with Low for the first two to four weeks of live trading on your broker, then stepping up once you've seen the EA behave on your actual fills and spreads. Going straight to High on a small account is a fast way to lose money in a bad week.

Symbol scope

XAUUSD only. The models were trained on gold-specific features and pip arithmetic. Running them on EURUSD, indices or crypto produces nonsense. Don't try it.

Updates

I plan to re-train the models every one to three months on rolling windows and ship updates via the Market auto-update. The four magic numbers, input layout and ONNX feature contract stay backward-compatible, so existing chart attachments keep working through updates.

Risk warning

Trading FX and CFDs is risky. You can lose all your capital. The numbers in this listing are historical and depend on broker conditions; your results will differ. I don't provide investment advice and I'm not liable for trading losses. Demo-test on your broker before going live, and never run the EA on capital you can't afford to lose.

Version 1.0 — initial release.