Testing Market Predictability with CNNs and Return Sequences

Summary

This research summary proposes using a convolutional neural network (CNN) to distinguish structured return sequences from shuffled sequences as a way to investigate weak-form market efficiency. In a simulation, pattern segments are inserted into white noise to create structured data, then shuffled to form a comparison set. The summary reports that classification improves as the signal-to-noise ratio rises. It also describes using hidden-layer activations and class activation heatmaps to locate the segments the model detects.

Reported market comparisons vary by asset and frequency: the CNN reportedly fails to distinguish daily index returns but identifies intraday index returns, with weaker recognition at five-minute than one-minute frequency. Individual-stock and factor daily returns are also described as difficult to identify; commodity-futures minute returns are more detectable than index-futures minute returns. These findings suggest hypotheses about where patterns may exist, not proof of persistent trading profits. The summary supplies no detailed performance metrics, and its conclusions depend on the selected data, model, and testing design.

Key ideas

• The proposed efficiency test trains a CNN to distinguish structured return sequences from shuffled ones.
• In simulated data, higher signal-to-noise ratios are associated with better sequence recognition.
• Visualization methods are used to locate return segments that influence the CNN's classifications.
• The reported results differ across asset classes and sampling frequencies, and do not by themselves demonstrate profitable strategies.

Tags

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