Engineering Price Angle Features for Machine Learning Models

Summary

The article investigates whether price slope and its arctangent angle can improve machine learning forecasts. It notes that time-based slopes depend on how gaps such as weekends are represented, while its alternative ratio of price changes can become unstable around Doji candles or zero denominators. The experiment uses USDZAR minute data and compares models trained on OHLC inputs, engineered angle and slope features, or both.

Across twelve models, OHLC alone performed best for simple linear regression, while the angle features helped only some models. K-nearest neighbors improved by 20% with the added features, and its tuned model was exported for an Expert Advisor. The article also describes trading filters involving MACD and dollar-related indicators, but the reported evidence is limited to its sample and setup. It does not establish that the features generalize across markets or periods, and it flags missing and infinite values as practical issues.

Key ideas

• Time-based price slopes can vary depending on whether elapsed market closures are included.
• A slope formed from open and close price changes can be unstable when the denominator approaches zero.
• The study compares OHLC inputs with price angle features across twelve machine learning models.
• K-nearest neighbors improved by 20% with the added features, while other models did not consistently benefit.
• Feature usefulness depends on the model and requires validation beyond the reported sample.

Tags

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