Implicit Switching-Cost Regularization in Supervised Trading Signal Classification

Automated trading systems built on supervised learning optimize prediction accuracy while ignoring transaction costs. This leads to volatile signals that trigger excessive position changes and destroy profitability. We show that conditioning predictions on the model’s previous output introduces an implicit switching cost into supervised learning, changing the effective optimization problem without modifying the loss function itself. This decision-path dependence reduces position change frequency while maintaining directional accuracy. Using major currency pairs sampled at 15-minute intervals, we find that models conditioned on previous predictions exhibit statistically significant reductions in switching frequency across all tested assets, with no meaningful deterioration in classification accuracy. Transaction cost sensitivity analysis demonstrates that this stability advantage mitigates performance degradation under increasing frictions. The implicit emergence of switching cost aversion through architectural design, rather than explicit penalty terms, offers a computationally tractable method for building transaction-cost-aware trading systems within standard supervised learning frameworks. The analysis uses foreign exchange data; we see no reason why the mechanism would not generalize to other asset classes, though empirical verification remains a direction for future work.