FxMath_RFConsensus is the first EA to offer two AI engines in one package: the original Random Forest (100% pure MQL, no DLLs) and the new LightGBM + Q-Learning version powered by a cross-compiled C++ DLL — gradient boosting with adaptive thresholds that learn from every trade.
Two AI engines in one package. The original Random Forest (pure MQL) and the new LightGBM + Q-Learning version (gradient boosting + RL).
Original Random Forest (pure MQL, zero dependencies) and new LightGBM gradient boosting (via static C++ DLL). LightGBM uses gradient boosting — it builds trees sequentially, each one correcting the errors of the previous. This converges to higher accuracy than bagging (RF) with fewer trees. Both share the same feature pipeline and multi-TF voting. Choose RF for zero-DLL simplicity, or LGBM for maximum accuracy.
Tabular Q-learning that dynamically adjusts trading thresholds based on market conditions. 27 discrete states (win rate x volatility x consecutive losses) map to 3 actions: Conservative (tighter filters after losses), Normal (balanced), Aggressive (looser filters on high win rate). Epsilon decays automatically from 0.30 by 0.998x per trade — early exploration, later exploitation. The Q-table persists across sessions, so knowledge compounds over weeks of trading.
No manual optimization or parameter tweaking needed. The EA retrains its models every N bars (default 200) using the latest price data — it automatically adapts to changing volatility, trends, and market regimes. Unlike static EAs that require frequent re-optimization, this EA learns from live market data continuously. The LGBM+RL version adds a Q-Learning agent that adjusts buy/sell thresholds based on every trade outcome — the EA literally gets smarter the longer it runs. Set it and forget it.
Both platform versions for both RF and LGBM+RL. Works on any broker type: ECN, STP, Market Maker, FIFO-compliant brokers. Compatible with hedged and netted accounts. Supports any symbol — Forex majors, minors, exotics, indices, commodities, metals, crypto, stocks. Any timeframe — M1 to MN1. Any leverage, any deposit currency. No broker restrictions whatsoever.
19 original (RSI, MACD, BB, SMA, ATR, momentum, candle body) plus 9 new: volatility ratio, linreg slope, volume ratio, skew, kurtosis, Elder-Ray, CMO.
Three independent ML models run on three different timeframes simultaneously (default M1, M5, M15). Each model independently analyzes its own timeframe and outputs an UP probability with a confidence score. The EA then applies confidence-weighted voting: a high-confidence M5 signal carries more weight than a low-confidence M15 signal. A trade is only executed when at least 2 of 3 timeframes agree. This eliminates false signals that appear on a single timeframe due to market noise. The result: fewer but higher-quality trades with statistically stronger confluence.
Training labels use ATR-relative thresholds. In high volatility the threshold widens to avoid noise. In low volatility it tightens to capture real moves.
Q-table saved to disk on shutdown, reloaded on startup. The RL agent remembers what it learned across sessions — knowledge compounds over time.
ATR-based SL/TP automatically adapts to symbol volatility — wider stops during high vol, tighter during low vol. Built-in money management sizes positions as a percentage of equity risk. Trailing stop locks profits as price moves in your favor. Break-even moves SL to entry once a target is reached. Time filter restricts trading to specific hours. Spread filter prevents trading during high-spread events. All configurable to match your risk tolerance.
Both share the same feature pipeline and trading logic. The LGBM+RL version adds gradient boosting and reinforcement learning.
The combination of gradient boosting and reinforcement learning creates a self-improving trading system that adapts to any market condition.
LightGBM builds trees sequentially — each new tree corrects the errors of all previous trees. This gradient boosting approach converges to higher accuracy with fewer trees compared to Random Forest's parallel bagging. LightGBM also handles non-linear relationships in financial data better than linear models, and its leaf-wise tree growth finds complex patterns that depth-wise trees miss. Result: significantly higher prediction accuracy on the same feature set.
Traditional EAs use fixed thresholds that become obsolete when market regimes shift. The Q-Learning agent continuously monitors win rate, volatility, and consecutive losses — three dimensions that capture the current market "personality." When conditions become choppy, the agent selects Conservative mode (fewer trades, higher quality). During strong trends, it switches to Aggressive mode (captures more moves). This isn't a simple rule — it's a learned policy that optimizes for long-term cumulative reward.
Most EAs use 3-5 indicators. This EA extracts 28 carefully designed features spanning: momentum (RSI, CMO, returns), trend (linreg slope, MA distance, Elder-Ray), volatility (ATR ratio, bar strength, Bollinger %B), volume (tick volume ratio), distribution (rolling skew & kurtosis), and cross-asset signals (MACD, MA crossovers). Each feature captures a different market dimension — together they form a complete picture of price action that a single indicator cannot provide.
The EA doesn't just predict "price up or down" — it uses ATR-relative labels. A bar is labeled UP only if future price moves by more than 0.3x the current ATR. In high volatility, this threshold widens (avoid labeling random noise as signal). In low volatility, it tightens (captures real directional moves). This makes the model robust across symbols and timeframes — no need to relabel for Gold vs EURUSD vs US30. The model learns the same underlying structure regardless of absolute price or volatility level.
Financial markets have multi-scale structure — a trend on M15 may be a pullback on H1. Running three independent models on different timeframes captures this hierarchy. The confidence-weighted voting system ensures that only signals with multi-timeframe confluence generate trades. A breakout that appears on M1 but not on M5/M15 is likely noise and is filtered out. When all three timeframes align, the probability of a genuine move is dramatically higher. The minimum vote threshold (default 2 of 3) is configurable.
Most EAs require constant optimization — run a backtest this week, parameters work. Next month, market changes, need to re-optimize. This EA solves that permanently. The model retrains automatically every N bars (default 200) using the most recent data. When volatility increases, it adapts. When trends emerge, it captures them. When ranging markets appear, it filters them. The RL agent adds another layer of adaptation, adjusting execution thresholds based on recent performance. This is truly a "set and forget" EA — it evolves with the market.
One package. Two versions. Lifetime license with free updates.