Neurons, activation functions, layers, forward propagation and gradient descent, applied to market prediction
|
Neural Networks - Introduction (pdf)
|
1.23 M |
The output of a neuron is a weighted sum z = w1 x1 + w2 x2 + b passed through an activation function, and the sigmoid gives a value between 0 and 1, read as the probability that a price rises. A single neuron has one linear decision boundary, a hidden layer is required for the XOR problem, and more hidden neurons give more curved boundaries. Sigmoid saturation gives tiny gradients, and ReLU is the common alternative. In forward propagation, market features are passed through input, hidden and output layers, and weights are learned by gradient descent on the loss landscape, with mean squared error for regression and cross-entropy for classification. In a case study on 100 test days, accuracy is 70 per cent, precision 69 per cent, recall 72 per cent and F1 0.71.