Neural Network

Neural networks are the architectural foundation of most "AI" EAs in the 2026 market. The terminology bears clarification because the marketing language outpaces the engineering reality — many products labelled "AI" use simple feed-forward networks of modest depth rather than the complex architectures (transformers, recurrent networks, deep convolutional networks) that the AI hype suggests.

Neural network basics for the EA context:

• **Architecture**: input layer (raw features from price data — current price relative to moving averages, ATR percentiles, volume ratios, etc.) → hidden layers (where the learning happens) → output layer (a single prediction: trade-quality score, direction probability, etc.) • **Activation functions**: introduce non-linearity that lets the network model relationships more complex than weighted sums; common choices are ReLU (rectified linear unit) for hidden layers and sigmoid for binary classification outputs • **Training**: backpropagation with gradient descent; the network's weights are adjusted to minimise prediction error on a training set of historical examples labelled as "trade succeeded" or "trade failed" • **Network sizes typical for EA applications**: input layer 10-30 features, 2-3 hidden layers of 50-200 neurons each, output layer 1 (probability of success) or 3 (probability of long, short, or no trade). Total parameter counts are 5,000-50,000 — modest by AI standards but adequate for the EA classification problem

How neural networks are used in modern EAs:

• **Signal filtering**: the most common architecture. A rule-based strategy generates candidate signals; the neural network scores each candidate and rejects those below a probability threshold. The net effect is higher win-rate without changing underlying strategy logic. Scalperology AI's M1 XAUUSD scalper uses this pattern • **Regime classification**: the neural network classifies current market structure as trending, ranging, or volatile, and the EA enables different sub-strategies based on the classification. Phalanx Neural AI's "abstain when uncertain" mechanism is an example • **End-to-end ML strategies**: the neural network IS the strategy — there's no separate rule layer. Inputs are price features; output is a trade signal directly. GoldStrike AI uses end-to-end ML with weekly retraining • **Position-sizing adaptation**: the network outputs a confidence score that adjusts position size — higher confidence trades get larger positions. Smart Robot AI's adaptive risk scaling uses this

Limitations and risks specific to neural networks in EA contexts:

• **Concept drift**: the relationships the network learned from historical data can change as market microstructure evolves. Without periodic retraining, networks decay • **Overfitting risk**: neural networks have many parameters and can fit historical noise rather than genuine patterns. Cross-validation discipline and out-of-sample testing are essential • **Black-box explanation**: unlike rule-based systems, neural network decisions can be hard to explain. "Why did the EA take this trade?" may not have a satisfying answer beyond "the network's output exceeded the threshold" • **Training data quality**: the network can only learn from the patterns in its training data; under-represented regimes will be poorly modelled. Markets that occurred only briefly during training may not be handled well live

For EA buyer evaluation, vendors using neural networks should document: (a) the architecture (layer counts, neuron counts), (b) the training data (period, instrument coverage), (c) the retraining cadence (weekly, monthly, quarterly), (d) the validation methodology (cross-validation, walk-forward, holdout). Vendors who don't document these are either hiding implementation details or hand-waving on an "AI" claim they don't actually engineer to.