Mitigating Transaction-cost Degradation in Financial Forecasting via Volatility-gated Inference

Deep neural forecasters for financial time series are conventionally optimized to minimize statistical loss functions such as the Mean Squared Error (MSE), yet they frequently suffer severe performance degradation when their raw predictive outputs are translated into live trading decisions. The principal culprit is not model inaccuracy per se, but the mismatch between the loss surface being minimized and the cost-laden objective that actually governs realized profit: bid-ask spreads, slippage, broker commissions, and the compounding drag of excessive portfolio turnover. In this paper we introduce the Volatility-Gated Inference (VGI) mechanism, a model-agnostic post-hoc policy that dynamically abstains from trade execution whenever normalized local volatility exceeds a calibrated threshold tau. We evaluate the proposal on EUR/USD daily exchange data from January 2019 through August 2026, comprising 1,980 trading days partitioned chronologically into 1,584 training and 396 out-of-sample testing days. The baseline directional model achieves a test MSE of 2.00 x 10^-5, yet its corresponding unconstrained trading strategy yields a cumulative return of -1.46% and an annualized Sharpe ratio of -0.13, evidence of capital erosion driven by over-trading. By contrast, the volatility-gated strategy, parameterized by a per-trade friction of c = 0.0001 and a normalized volatility threshold of tau = 1.0 sigma, achieves a cumulative return of +3.46% and an annualized Sharpe ratio of +0.36, while simultaneously reducing trade switches from approximately 280 to 145. These results demonstrate that selectively abstaining from execution in volatile regimes can invert a negative-expectancy strategy into a profitable one without any modification to the underlying predictive model.