okhp3-outcome-modeling-core

Aggregate raw events into a time-indexed entity state using only information available at the as-of time, then estimate probabilities with a transparent model such as logistic regression. Validate with time-based splits and calibration metrics like Brier score and log loss rather than random row splits.

Build a feature inventory recording source, grain, cadence, missingness, and leakage risk, then test nested tiers such as 10, 30, 60, 120, and 200 features. Compare each tier against a simple baseline using time-aware out-of-sample metrics and stop when incremental value is negligible.

A forecast estimates the probability of an outcome, while a decision applies an explicit objective function with costs, capacity limits, and risk constraints. This Skill keeps the two separate so the same evidence can support forecasting, ranking, budgeting, or market comparison.

No external dependencies are required. The method works as portable prose, and the optional helper script uses only the Python 3.9+ standard library to compute weighted means, logistic probabilities, Brier scores, and expected values from local JSON.

Do not use it to invent data, current facts, or probabilities, to claim causation from correlation without a causal design, or to execute trades or bets. When domain rules matter, load a narrow domain adapter instead of relying on the generic core.

A model can rank well yet be poorly calibrated, or forecast accurately while producing a bad allocation under real costs and constraints. Check calibration, subgroup performance, drift, and the decision objective separately rather than relying on accuracy alone.