Normalizing Time-Series Inputs for Neural Network Trading Models
Summary
The article explains why neural networks used in trading may benefit from scaling their inputs. It introduces z-score normalization, which centers values around their mean and scales by standard deviation, and min-max normalization, which maps values using the sample minimum and maximum. It also notes that target values can be normalized in regression tasks. The practical example applies min-max scaling to indicator data for a perceptron-based Expert Advisor, and the article describes comparisons between versions using raw and normalized inputs.
The discussion cautions that min-max scaling can be distorted by outliers, changing data ranges, or unrepresentative samples; local scaling, z-scores, smoothing, or outlier treatment may be alternatives. The article does not provide quantitative performance results in the supplied text, so it does not establish that normalization improved profitability or forecasts. The suitable method depends on the data and task, and normalization alone does not ensure a reliable trading model.
Key ideas
- Input normalization can improve neural-network training stability and convergence by bringing features onto comparable scales.
- Z-score normalization centers data and scales it by its standard deviation, while min-max normalization scales it relative to sample extremes.
- Outliers and changing time-series ranges can make min-max scaling misleading, so alternative or local preprocessing may be appropriate.
- The article compares perceptron EA setups using raw indicator values and normalized values, but the supplied text gives no quantitative comparison results.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.