Market Trader AI Pro

Market Trader AI Pro

• EURUSD
• EURCHF
• EURGBP
• USDCAD
• AUDCHF
• AUDCAD
• AUDUSD
• USDCHF
• GBPJPY
• AUDJPY
• GBPUSD
• EURAUD
• EURJPY
• USDJPY
• CHFJPY
• XAUUSD
• NZDUSD
• NZDJPY
• NZDCHF
• NZDCAD
• GBPNZD
• GBPCAD
• GBPAUD
• CADJPY
• CADCHF
• USDSGD
• EURCAD
• AUDNZD
• GBPCHF
• EURNZD

Technical Strategy Description – Market Trader AI Pro 

Market Trader AI Pro v1.03 is a fully automated, self-adapting Expert Advisor for MetaTrader 5 that combines classical price-action logic with an embedded, continuously retraining neural network. The robot works exclusively on the H1 timeframe and currently supports 28 major and cross currency pairs.

Core Architecture

• Neural Network Topology: 112 → 7 → 1 fully connected feed-forward network
• Input Layer (112 neurons): Normalized closing prices of the last 112 bars (approximately 4.5 days on H1). Each price is normalized individually per symbol using a running min-max scaler: normalized_price = 2 × (price − min) / (max − min) − 1 → range [-1, 1]
• Hidden Layer (7 neurons) + Output Layer (1 neuron): hyperbolic tangent (tanh) activation function both for hidden and output layers.
• Activation derivative: 1 − x² (used during back-propagation).

Pre-training & Embedded Weights

• For each of the 28 supported symbols, a separate set of optimally trained weights and biases is compiled directly into the .ex5 file as binary resources (#resource "weights_1.03XXXYYY.bin").
• At startup the EA automatically detects the current chart symbol and loads the corresponding pre-trained weights — no generic “one-size-fits-all” model.
• If (for any reason) the resource is missing, the network initializes with small random weights and immediately starts training from scratch.

Online Self-Retraining Mechanism

• Retraining trigger: every 20 new completed H1 bars (input parameter TrainAfterBars, default = 20).
• Training dataset: the most recent 340 bars (input TrainingBars).
• Target calculation: simple binary direction of the next bar target = (Close[i-1] > Close[i]) ? +1 : -1
• Training process: 270 epochs (input Epochs) of standard stochastic gradient descent with fixed learning rate 0.0155.
• After retraining is completed, the updated weights instantly replace the current ones — the robot literally “learns” while trading live.

Signal Generation

• On every new H1 bar the network performs a forward pass using the latest 112 normalized closing prices.
• Raw output is interpreted as follows: Buy → output > +SignalThreshold (default 0.0005) Sell → output < −SignalThreshold
• Only one position per symbol is allowed at any time (no grid/martingale).

Money Management

• Risk per trade: 0.2 % of current balance by default (input RiskPercentage = 0.002).
• Lot size is calculated dynamically using tick value and a very wide protective Stop-Loss (50 000 points = 5000 pips on 5-digit quoting) — effectively the SL is almost never hit; it exists only for broker compliance and lot-size calculation accuracy.

Take-Profit & Exit Rules

• Fixed Take-Profit = 70 points (7 pips on 5-digit quotes) by default.
• No trailing stop, no breakeven — the strategy relies entirely on the neural network’s short-term directional accuracy.

Key Advantages of This Approach

1. Extremely low prediction horizon (1 bar ahead) → the network only has to be slightly better than 50/50 to generate consistent profit.
2. Per-symbol specialized weights + continuous online adaptation → the EA automatically adjusts to regime changes (trending ↔ ranging ↔ high-volatility periods).
3. No external dependencies — everything (weights, training, inference) runs inside pure MQL5, works on VPS without Python or DLLs in the current version.