Forecasting Conditional Price Distributions with an MLP and Gaussian Likelihood

Summary

This article describes a multilayer perceptron that forecasts a conditional Gaussian distribution for price increments rather than producing only a point estimate. It motivates the approach by noting that mean squared error corresponds to a Gaussian assumption with constant variance, while financial return variability can change with the inputs. The model therefore predicts both a conditional mean and variance, using Gaussian negative log-likelihood as its training objective. A positive-output activation constrains the variance estimate, and the implementation uses L-BFGS optimization.

The article walks through the likelihood rationale, model structure, and an indicator application that displays a mean forecast with confidence intervals for a currency pair. A synthetic normal sample illustrates maximum likelihood estimation of a mean, while the indicator demonstrates the intended use on market data. These are demonstrations rather than a rigorous comparison of forecasting performance: the supplied text gives no out-of-sample results or calibration analysis. The Gaussian distribution assumption also limits the model’s representation of skewness and heavy tails, so predicted intervals should not be treated as guaranteed risk bounds.

Key ideas

• Mean squared error implicitly fits a constant-variance Gaussian regression model.
• Gaussian negative log-likelihood lets an MLP predict conditional means and variances together.
• A positive activation is used for the variance output, while the mean output is linear.
• The indicator presents forecasts with confidence intervals rather than point estimates alone.
• The Gaussian assumption and lack of reported out-of-sample evaluation limit the conclusions.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.