Constructing neural trading systems from facts to walk-forward

Specify the procedure first

A neural trading procedure is specified by five construction choices: the forecast object, the input requirements, the network type, and the system's size and structure. Those choices decide what the network is asked to emit, which series it may see, how layers are arranged, and how large the system is.

The procedure can be configured to emit one of four output kinds: classification, pattern, real-valued levels, or an optimization result.

Build the fact schema

Inputs may mix technical and fundamental series. Apparently peripheral series may still be kept, because training is expected to sort relevance. Preprocessing those inputs into relative forms such as oscillators, momentum, and ratios is treated as necessary so relationships are explicit rather than left for the network to invent.

A fact is a paired record of current market and rate inputs plus the later target the procedure is asked to map. Input-layer width equals the number of data categories after preprocessing. A twelve-category market-and-rate set implies twelve input neurons, and each paired input with its next-session target is a training fact.

Set size and structure

A typical feedforward layout uses an input layer, a hidden layer, and an output layer. Information moves only to the next layer. Neurons in the same layer do not connect. A fully connected feedforward layout joins every neuron to every neuron in the succeeding layer.

The hidden layer is the interior recoding stage meant to hold correlations that are not explicit in the raw inputs. Hidden-layer capacity is a generalization control. Too few neurons block input-to-output mapping. Too many let the network memorize training facts instead of extracting features, which is treated as equivalent to curve-fitting or overoptimization.

Output-layer width equals the number of forecasted categories. Forecasting the next session's open, high, low, and close requires four output neurons.

Train with a learning law

Supervised training presents facts, then uses a learning law such as backpropagation to send error backward and change connection weights only when the forecast mismatches the target. Once all facts are learned, the network is treated as stable and ready for further testing.

Hold weights fixed in walk-forward testing

Walk-forward testing uses facts withheld from training and does not update connection strengths. Live use likewise feeds daily inputs without weight changes. A trading fact file is a live input file that matches the training schema but omits the target patterns present in the training file.