Neural Network
A machine-learning model made of interconnected layers of simple computing units (neurons) that learn to map inputs to outputs by adjusting weighted connections during training.
A neural network is a family of machine-learning models loosely inspired by how biological neurons connect. It is built from layers of simple units, each of which takes a set of numeric inputs, multiplies them by learned weights, adds a bias, and passes the result through a nonlinear activation function. Stacking these layers lets the network compose simple transformations into complex ones, so it can approximate relationships that linear models cannot capture. Training works by feeding in examples, comparing the network's output to the correct answer with a loss function, and using backpropagation and gradient descent to nudge the weights in the direction that reduces error.
The reason neural networks matter is that they are universal function approximators: given enough data and capacity, they can learn intricate, nonlinear patterns without a human specifying the rules in advance. In finance this is appealing because asset returns depend on many interacting factors in ways that are noisy and far from linear. The same flexibility is also the main hazard. A network with too many parameters relative to the available data will memorize noise rather than signal, a failure called overfitting, which is why honest validation, regularization, and out-of-sample testing matter more here than raw model size.
Plain feed-forward networks treat each input independently and have no sense of time, which is a limitation for markets where order and history matter. That gap is why specialized architectures exist, such as recurrent networks, temporal convolutional networks, and attention-based transformers, each of which adds structure for handling sequences. A neural network is best thought of as the general building block; the architecture choice determines how that block is wired to suit the data.
Hedgewing.ai is built on neural networks rather than simple linear scoring. Its forecasting engine is an ensemble of four deep-learning architectures (LSTM, GRU, TCN, and Transformer) that each read the same set of 45 engineered features and learn complementary patterns. Because neural networks can overfit, hedgewing.ai validates every model with nightly walk-forward backtesting and reports calibrated confidence alongside each prediction, so the flexibility of the models is paired with a discipline for measuring whether they actually generalize.
Related terms
Deep Learning · Ensemble Model · Overfitting · Feature Engineering
Back to the hedgewing.ai glossary · See AI stock forecasts