A good trade score can still produce a poor portfolio

Scoring trades independently can reward decisions that create an undesirable portfolio together. We propose comparing an isolated-trade objective with a declared portfolio objective on the same saved opportunity sequences. The experiment remains unrun. It asks whether evaluation changes behavior in a way that matches the owner's aggregate constraints.

The continuous record companion reports historical positions, liquidation concentration and configuration effects. Those findings motivate examining exposures jointly, while providing no result for a new portfolio objective. The controls paper companion supplies the authenticated mandate and reconciled state needed to know which portfolio the agent is managing.

Two individually admissible entries

An illustrative account has $10,000 equity. Two proposed positions each add $2,000 notional, and each satisfies a $2,500 single-position cap. If they are strongly exposed to the same market factor, their combined $4,000 exposure may violate a separately declared $3,000 factor limit. A grader that sees only the two trade records can award both full credit while the portfolio violates the owner's objective.

Use a simple stress calculation before introducing an estimated correlation matrix. Suppose a specified factor shock loses 5% on each position. The losses are $100 and $100, so the simultaneous portfolio loss is $200, or 2% of equity. The arithmetic illustrates aggregation under a chosen shock; it does not estimate the probability of that shock. More elaborate risk models should declare their estimation window and uncertainty.

Create episode-level inputs that include existing positions, pending orders and available margin. Keep action opportunities identical between objectives. One arm grades forecast quality and isolated trade outcome. Another applies a predeclared portfolio utility with exposure constraints, turnover cost and an explicit treatment of missed opportunities. Both retain common deterministic policy checks.

Evaluate the whole sequence

Report portfolio drawdown, risk-limit violations, turnover and costs over complete episodes. Preserve trade-level scores as diagnostics, so a weaker forecast channel cannot disappear inside a favorable aggregate. Include a no-action reference and a simple allocation policy. An objective that merely suppresses all entries can reduce losses while failing the owner's opportunity mandate.

We would audit whether pending orders are included when calculating exposure. Two agents acting on the same account snapshot may each satisfy a cap individually and exceed it together. That case tests state reservation and reconciliation rather than forecasting. It belongs in the episode even if the final trades happen to make money.

The proposed contribution is an explicit mapping from owner objective to evaluation unit. If the owner cares about the portfolio, the primary outcome should be attached to the portfolio sequence. Isolated trades remain useful for diagnosing information and execution errors. Choosing between objectives would require measured behavior, complete cost accounting and uncertainty over independent market episodes, with any conclusion scoped to the objective actually tested.