Testing Higher Highs and Lower Lows with Machine Learning

Summary

This article compares forecasting price direction with predicting whether a future close will exceed the current high, fall below the current low, or remain between those levels. The latter target formalizes a classic price action idea: traders look for successive extremes as signs of a developing trend, while failures to extend may indicate weakening momentum. The article also describes the difficulty of acting on a breakout when retracements can quickly reverse it.

The authors train several classifiers, including decision trees, AdaBoost, and neural networks, and compare them using time-series cross-validation without shuffling. Exploratory plots show that current close, high, and low do not clearly separate the target classes; the three-state target also leaves many ambiguous cases. The article reports that simpler price-change forecasts may be more effective than the higher-high/lower-low target. Its evidence is limited to the described dataset and setup: models were not tuned, and the authors suggest that larger datasets could provide more insight. The classification labels and examples also contain apparent inconsistencies, so the implementation details warrant careful checking before reuse.

Key ideas

• The study compares a simple future price-direction target with a three-state target based on the current high and low.
• Price action traders use successive extremes to identify possible trend formation and failed extensions to assess weakening trends.
• The plotted features show little natural separation between the price-action target classes.
• The article reports that simpler forecasts of price changes may work better than the more complex target.
• Time-series cross-validation preserves temporal ordering, but the reported analysis is limited by its dataset and untuned models.

Tags

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