" the Hourly subset of the M4 dataset for public cross-scenario validation"

"Selecting useful predictive objects in dynamic environments is fundamentally different from evaluating them with static performance indicators. Existing practices often rely on current-level statistics, such as return, Sharpe ratio, drawdown, turnover, or prediction error, to rank candidate objects. However, for evolving objects such as quantitative factors, signals, forecasting series, or predictive models, future utility depends not only on current quality, but also on temporal evolution, structural efficiency, and lifecycle state.In this paper, we propose a Future Capability Meta-Feature Framework (FCMF) for modeling the future capability of evolving predictive objects. A key idea of the framework is to describepredictive objects in a way that is closer to how evolving systems are characterized in the physical world: future behavior should not be inferred from current level alone, but from state, state change, efficiency structure, and nonlinear response. Instead of treating candidate objects as static entities, FCMF represents each object through operatorized meta-features derived from its current state. These meta-features are constructed using three operator families: difference operators for temporal changes and decay\/rebound patterns, ratio operators for efficiency and normalized trade-off characterization, and transform operators for nonlinear state adjustment and structural reshaping. The resulting representation enables future utility prediction and dynamic ranking under rolling evaluation settings. We instantiate the framework in a large-scale quantitative factor selection problem, where the goal is to predict the future out-of-sample utility of factors from their current state descriptors. To examine cross-scenario generality, we further validate the framework on a public time-series benchmark by recasting time series as evolving predictive objects and predicting their future forecastability. Results indicate that operatorized meta-features provide stable predictive signals for future utility ranking, bucket separation, and Top-N selection over static baselines and simpler rolling models. These findings suggest that predictive object selection should be formulated as a dynamic future capability modeling problem rather than a static score ranking problem. More broadly, theproposed framework provides a general data-centric approach for representing evolving objects through future-relevant state abstractions and for supporting future-aware ranking and selection in non-stationary environments."