Self-Supervised Targets from OHLC Midpoints for Trading Models

Summary

The article proposes generating alternative classification targets directly from OHLC observations instead of relying only on future closing price. It calculates pairwise price midpoints, labels whether each candidate value rises over a chosen forecast horizon, and compares targets using the same linear discriminant model and time-series cross-validation. The intended benefit is to find input-derived targets that may be easier for a model to learn, addressing the risk that standard supervised targets do not have a learnable relationship with the available observations.

The worked example uses broker-supplied EURUSD data and reports that model accuracy did not translate directly into profitability; the stated accuracy was 68% while profitability was 52%. The article also describes a longer unsupervised run and cautions against assuming that adding more data improves accuracy. These are results from one example and model setup, not evidence of general performance across instruments or market regimes. Statistical accuracy and trading profitability are distinct measures, and the material does not establish that the proposed targets will remain useful out of sample.

Key ideas

• OHLC values can be transformed into multiple candidate future targets for a self-supervised learning exercise.
• The example compares candidate targets with the same model and time-ordered cross-validation.
• The author evaluates binary labels based on whether a future candidate value exceeds its current value.
• The reported classification accuracy exceeds the reported profitability, illustrating that the measures differ.
• The EURUSD example does not establish that the approach generalizes to other markets or regimes.

Tags

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