FxMath Trading AI — Documentation

Overview

Both versions are self-contained Expert Advisors that train machine learning models directly inside MetaTrader. They collect price data, extract technical features, train models, and combine predictions from three timeframes via confidence-weighted voting.

Version Comparison

Installation

Both Versions

1. Copy .mq4 / .mq5 to MQL4\Experts\ or MQL5\Experts\
2. Compile in MetaEditor (F7)
3. Attach to a chart

LGBM+RL Only — DLL Setup

1. Copy LGBM_RL_Engine.dll to MQL4\Libraries\ or MQL5\Libraries\
2. Restart MetaTrader or use Tools → Options → Expert Advisors: Allow DLL Imports
3. The DLL is fully static — only depends on KERNEL32, msvcrt, WS2_32, IPHLPAPI

ML Engines

Random Forest (v1.1)

Ensemble of decision trees trained with bootstrap aggregation (80% sample) and Gini impurity splitting. Hyperparameters: InpNumTrees (5-50), InpMaxDepth (3-15), InpMinSamples (2-50).

LightGBM (v1.0)

Gradient boosting via the C API of Microsoft's LightGBM library. Hyperparameters: InpNumLeaves (15-127), InpLearningRate (0.01-0.3), InpNumIterations, InpMinDataInLeaf, InpFeatureFraction, InpBaggingFraction. Uses objective=binary with metric=binary_logloss.

Multi-Timeframe Voting

Three independent models on three configurable timeframes. Each predicts UP probability. The EA combines them:

• BUY when at least InpVoteMin TFs have probability above BuyThreshold
• SELL when at least InpVoteMin TFs have probability below SellThreshold
• NEUTRAL otherwise

Votes are weighted by confidence (|prob - 0.5| × 2). In the LGBM+RL version, BuyThreshold, SellThreshold, and MinConfidence are dynamically adjusted by the RL agent.

28 Features (LGBM+RL) / 19 Features (RF)

All features are computed per-TF and normalized. The RF version uses features 1-19. The LGBM+RL version adds 9 more.

Original 19 Features (both versions)

New Features 20-28 (LGBM+RL only)

Q-Learning RL Agent (LGBM+RL Only)

The Q-Learning agent dynamically adjusts trading thresholds based on market conditions. It uses a tabular Q-table with 27 states × 3 actions.

State Space

Default total: 27 states. The state is computed each new bar before prediction.

Actions

Reward

reward = profit / ATR_at_open — normalized to volatility. Clipped to [-3, +3].

Learning Loop

1. On each new bar: discretize market state → get action via epsilon-greedy
2. Apply action adjustments to thresholds
3. Predict, vote, trade
4. When trade closes: compute reward → update Q-table → decay epsilon

Q-Table Persistence

The Q-table is saved to InpRLSavePath (default: RFConsensus_RL.qtable) on OnDeinit and loaded automatically on OnInit. Knowledge accumulates across sessions.

Training Pipeline

1. Data Collection: Fetch InpTrainBars (default 500) of OHLCV data per TF
2. Feature Extraction: Compute all 19/28 features for each bar
3. Label Generation: Label = 1 if future close > current close + 0.3 × ATR(14), else 0
4. Training: Train RF (in-memory) or LGBM (via DLL) on feature matrix
5. Retrain: Every InpRFRetrainBars / InpLGBMRetrainBars bars

Labels are volatility-adaptive — in high-volatility regimes, the threshold widens so the model focuses on meaningful moves, not market noise.

Trading Logic

At most one position at a time. When a new signal arrives:

• No position → open trade at market
• Has position → let SL/TP close it (no reversal — avoids overtrading)

Filters: spread check, time filter (InpTimeFilterEnable), minimum confidence.

SL/TP from ATR on main TF. Trailing stop and break-even update every tick.

RF v1.1 Parameters

LGBM+RL v1.0 Parameters

RL Parameters (LGBM+RL only)

Common Parameters (both versions)

Chart Panel

When InpShowOnChart = true, an information panel appears on the chart:

• Title — EA name and version
• VOTE result — BUY/SELL/NEUTRAL with agreement count
• 3 × TF rows — probability and confidence per timeframe
• Entry / SL / TP levels
• RL Status (LGBM+RL only) — action, epsilon, trade count
• Position status — direction and floating profit

FxMath_RFConsensus v1.1 · FxMath_RFConsensus_LGBM_RL v1.0 · Documentation