Product Launch
This month we started selling our newly developed EA — DeepAgent FX — on the MQL5 Market and GogoJungle. No Martingale, no hedging: a next-generation FX AI agent trained on large-scale market data.
What sets DeepAgent FX apart from thousands of other EAs? AI does the filtering. Every candidate trade from our trend engine must pass an ensemble of machine learning models before execution. Below is what that filter actually changes in practice.
ML Filter vs Baseline: Why Continuous Learning Matters on XAUUSD
DeepAgent FX is now live on real accounts. The AI ensemble filter sits between our trend engine and execution — rejecting low-quality setups and keeping only trades the models believe in. Below: live results, then why the filter matters.
MT5 mobile trade history — XAUUSD, Jun 22–26, 2026
Results shown are from our internal research build, which retrains the AI model every ~3 days. The marketplace version ships with a fixed model; see our forward-test panel for ongoing live performance.
| Metric | With AI filter | Without AI filter | Δ |
|---|---|---|---|
| Net Profit | $7,054.65 | $2,383.13 | +196% |
| Total Trades | 316 | 587 | −46% |
| Profit Factor | 1.73 | 1.09 | +59% |
| Expected Payoff | 22.26 | 4.06 | +448% |
| Drawdown % | 21.77% | 32.52% | −33% |
| Recovery Factor | 5.55 | 0.58 | +857% |
| Sharpe Ratio | 15.81 | 2.35 | +573% |
Six Stages: From Data to Live Trade
Millions of M5 bars + tick data. Price, trend, volatility, regime, volume.
EMA pullback rules — the baseline strategy (no ML yet).
Random Forest, XGBoost, CatBoost, Logistic Regression + calibrated consensus.
Side-specific thresholds (e.g. Long 0.32 / Short 0.28). Reject weak signals.
Lot multiplier scales with model confidence (1x–3x).
ONNX models embedded in EA. Trailing stops + session risk rules.
A static ML model is a snapshot of the past. Gold regimes shift — volatility clusters change, session dynamics evolve, macro events rewrite correlation structures. The baseline strategy (Pass 0) has no mechanism to adapt; it fires every rule-based signal regardless of current market quality.
Our Continuous Learning Loop
- Batch backtest — collect every trade + 20 features at entry
- Walk-forward selection — pick best model on out-of-sample 30%
- Full-data refit — retrain deployment copy on 100% recent window (aligned Python + MT5 ONNX)
- Side-specific deploy — separate Long/Short models + thresholds to MT5 EA
- Performance feedback — live results feed next batch cycle (target: retrain every ~3 days)
Key insight
The ML advantage in Pass 1 is not magic — it is learned selectivity from recent trade outcomes. Without continuous retraining, that edge decays. With it, the filter stays calibrated to the market you are actually trading today.
- ML does not replace the trend strategy — it filters it.
- Same rules, 3× profit, half the trades, one-third less drawdown.
- Side-specific models (Long ≠ Short) outperform a single common threshold.
- Continuous learning closes the loop: live → data → retrain → redeploy → live.
- Python live filter and MT5 ONNX EA share identical decision logic.