Neural networks

A neural network is a stack of simple units ('neurons') arranged in layers. Each unit computes a weighted sum of its inputs, passes it through a nonlinear function, and feeds the result forward. With enough units and layers, a network can approximate astonishingly complex relationships — this is deep learning.

How it learns

Training adjusts the connection weights to reduce prediction error, using backpropagation (computing how each weight contributed to the error) and gradient descent (nudging each weight in the direction that lowers error). Repeat over millions of examples and the network gradually shapes itself to the data.

The promise for markets

Neural nets can capture nonlinear patterns and interactions that linear models miss entirely. Given enough data they can blend price, volume, fundamentals and alternative data into a single prediction without a human specifying the relationships.

The dangers, stated plainly

• Data hunger. Deep nets need vast amounts of data. Financial history is short and largely non-repeating — there is only one 2008. This scarcity is the deepest problem in applying deep learning to markets.
• Overfitting. A large network has millions of parameters and will memorise noise unless heavily regularised.
• Opacity. A neural net is a black box; understanding why it predicts something is hard, which makes failures hard to anticipate.
• Non-stationarity. A net trained on one regime can fail silently when the market changes character.

Neural networks are genuinely powerful and genuinely dangerous in finance. They reward teams with abundant data, strong validation discipline and humility — and punish everyone else. The next lesson covers the architecture built specifically for sequences.