This is a price-based FX trend system built as a dual-horizon exposure controller (mean hold ~1h, P95 ~11h) with auto-hedging between short- and long-horizon risk. The design enforces a two-edge constraint: (1) a mode edge that targets positive expectancy per unit time through short-duration trades, and (2) a tail edge that captures rare, large directional moves by selectively extending holding time when continuation remains justified; otherwise the trade is closed quickly to control time-in-market risk.

The strategy fuses two independent trend-state estimators that share the same control objective (hold vs release) but use different representations of price dynamics to reduce regime sensitivity and edge decay. One estimator models the trend lifecycle (birth/death likelihood, persistence/memory, and acceleration-style state changes). The second performs a Trend/Cycle/Residual decomposition and forecasts each component using a heterogeneous toolkit (including Kalman-based epistemic filtering under fuzziness / Possibility Theory, plus ML/DL and quantile regression where appropriate). Outputs are combined via entropy-based fusion that explicitly accounts for epistemic uncertainty (model confidence / ambiguity) and aleatoric market noise, with the weighting chosen for stability and robustness rather than maximizing a single backtest metric.

Risk is handled with explicit structural constraints: the system does not flip positions (only flatten/hold), avoiding "self-correction by reversal" in noisy conditions. When positions are promoted to the long-horizon regime, the strategy can scale/overlay exposures primarily to hedge newer trades and stabilize aggregate PnL. Validation is performed with cross-validation and overfitting control using De Prado's Probabilistic Backtest Overfitting (PBO); performance is assessed separately on mode vs tail segments to ensure both edges contribute independently rather than one being an artifact of the other.