Quant Scavenger Gold Fbs
- Uzmanlar
-
Frank Michel Noughue Lemoupa
Software Engineer and algorithmic trading developer specializing in data-driven Expert Advisors for MetaTrader 5. My focus is on building reliable, well-tested trading systems backed by research and continuous improvement.
Explore my products: - Sürüm: 1.0
- Etkinleştirmeler: 20
Overview
Quant Scavenger – Gold Series EA is part of a structured series of single-asset, ML-based quantitative Expert Advisors.
Each EA in the series is dedicated to one broker-offered asset only. This edition is specifically built for trading Gold (XAUUSD CFD).
The embedded quantitative model is trained exclusively on H1 (1-hour timeframe) data. It is not a multi-timeframe system and is designed to operate on H1 only.
Core Design Principles
Single asset, single timeframe specialization An ensemble of four independently trained model types — Random Forest, XGBoost, Gradient Boosting Machine, and LightGBM — vote on trade direction. A trade only fires when a configurable minimum number of models agree, reducing single-model overfit risk.
A Follow-the-Leader style selection policy (LazyFTL) periodically re-scores which model combination has performed best on recent closed trades (by win rate or other configurable reward metrics) and can reallocate to a different combination live — the system is not statically frozen at purchase time. Pair/model candidates are evaluated from a seed-shuffled subset pool, with a configurable minimum score a candidate must beat before it's adopted.
Stop-loss distance is derived from MAE/MFE analysis on historical trades (how far winning and losing trades actually moved before resolving), rather than an arbitrary fixed-multiple stop.
Portfolio-based risk allocation
Structured, non-scalping execution
This EA is not a generalized multi-market robot. It is engineered specifically for Gold CFD price structure.
Model & Updates
The quantitative model:
- Is trained on historical H1 data using a walk-forward validation methodology, with evaluation designed to avoid the accuracy inflation that comes from overlapping test windows
- Uses structured feature inputs derived from price behavior (including ATR, RSI, Stochastic, Williams %R, and CCI)
- Is embedded into the EA logic for fast in-EA inference
- Is reviewed and retrained periodically
Monthly/yearly updates allow:
- Adaptation to evolving volatility regimes
- Recalibration to structural market changes
- Improved robustness over time
Each update reflects refreshed training and optimization based on recent market data.
Verified Performance
The performance are verified for FBS only, due to data availability.
Broker Consideration
Primary optimization performed using FBS CFD data.
Because quantitative systems are sensitive to tick data differences, spread structure, and execution model — and most CFD brokers are also market makers — backtesting on your own broker is recommended before live deployment.
Risk sizing (stop-loss/take-profit) is defined in fixed USD amounts rather than raw price points, so behavior stays consistent across brokers with different tick sizes. A USD-denominated account is assumed for the default risk settings; non-USD accounts should re-check position sizing before going live.
The MAE and MFE analysis advise 58 USD as the account-risk stop-loss basis.
Trading Conditions
- Required timeframe: H1 only
- Minimum deposit: 350$ advised; a cent account is a reasonable alternative if you're not ready to commit that amount
- Instrument: Gold CFD (XAUUSD)
Not designed for:
- M1/M5 scalping
- Multi-timeframe stacking
- Multi-asset trading inside one EA
Inputs
Risk Management
- max_lot — hard ceiling on position size regardless of other sizing settings
- base_lot — starting lot size for fixed-lot mode
- Take-profit / stop-loss — account risk expressed as a fixed USD amount per trade, broker-agnostic
Position Sizing
- Fix_Lot_Mode — if true, a constant lot size is used instead of dynamic sizing
- Fix_Lot — used only when Fix_Lot_Mode = true . If set to 0, the broker's minimum lot size is used
Model Configuration
- ilook_back — historical lookback window used by the model
- period — base indicator period
- UseLazyFTL — enable adaptive pair/model reallocation
- LazyFTLEvalWindow — how many closed trades are evaluated per reallocation cycle
- Active models — which of the four models (RF, XGB, GBM, LGBM) are active in the voting pool; full ensemble by default
- Min models that must agree to fire — minimum number of active models that must agree before a trade fires (0 = full consensus required)
Lazy-FTL Pair & Model Selection (all optional, off by default unless noted)
- Candidate pool bucketing (number of buckets, which bucket to use, shuffle seed) — controls how the candidate pool is subset for evaluation, for reproducibility
- Candidate must beat this absolute score / must beat active pair by — thresholds a candidate must clear before replacing the active configuration
- Enable model-selection lazy-FTL policy — turns on adaptive model switching
- LazyFTLModelReward — reward metric used to score candidates (e.g. win rate)
- LazyFTLModelArmMode — ensemble arm mode (e.g. Mode A: 15 ensemble)
Trial Version
Contact me if you have a broker in mind and want a free demo.
Operational Notes
- Not intended to guarantee immediate profits
- Not immune to market shocks or slippage
- Manual supervision during major news events is advisable
- Past performance, whether backtested or live, does not guarantee future results
Series Structure
The Quant Scavenger Series will expand to include one EA per asset, each trained and updated independently. This Gold Series edition follows the Apple Series as the second release in that structure.
Support
Contact me at frankmichel022@gmail.com — or https://www.mql5.com/en/users/frankmichel022-gmail, if you went through the trouble of downloading this, you might as well write the developer for suggestions.
