Neural Networks and Nonlinear Predictions
EVIDENCE LAB · 14
Neural Networks
& Nonlinear Prediction.
Neural networks construct flexible nonlinear functions from interconnected layers. Learn how architecture creates capacity—and why generalisation, transparency and economic judgement remain essential.
WEIGHTACTIVATELEARN
01 · FORWARD PROPAGATION
Layer by layer,
features become predictions.
Input layer
Receives standardised economic features available at prediction time.
Parameters
Weights and biases determine how signals combine across connections.
Activation
Nonlinear transformations such as ReLU allow flexible functional forms.
Output layer
Produces the continuous forecast, probability or classification score.
02 · LEARNING FROM ERROR
Optimise the loss.
Protect generalisation.
Loss function
L(θ)=mean(Y−Ŷ)²
Quantifies prediction error for the task being learned.
Gradient descent
θₜ₊₁=θₜ−η∇L(θₜ)
Updates parameters in the direction that reduces training loss.
Regularisation
L*=L+λΩ(θ)
Constrains capacity through penalties, dropout or early stopping.
NEURAL NETWORK ARCHITECTURE STUDIO
Shape the network.
Watch capacity respond.
Change depth, width, activation, learning rate, noise and regularisation. ECONORIA redraws the architecture and estimates training versus validation behaviour.
ARCHITECTURE VERDICT
03 · ECONOMIC DISCIPLINE
Flexible prediction does not replace
economic understanding.
Use temporal splits
Future observations must never leak into preprocessing or training.
Benchmark simply
A neural network must outperform transparent linear and tree models.
Interpret cautiously
Attribution methods approximate model behaviour, not causal mechanisms.
Monitor stability
Nonlinear models can fail abruptly when economic regimes change.
