Portfolio Agent EvalsHarbor task suite · ETF analyze → backtest → rebalance

Themes

Each theme isolates one capability. Tasks within a theme climb the tier ladder so a failure pattern is legible: a model that passes tier 2 engine tasks but fails tier 3 look-ahead audits tells you something specific.

01Analyze

Data Forensics & Canonicalization

Every downstream number is only as good as the price panel. These tasks hand the agent multi-vendor ETF price files with planted defects (format drift, duplicates, unit scaling, undocumented splits, mis-stamped dividends, calendar mismatches) and require a canonical, adjusted, calendar-aligned panel plus an explicit anomaly ledger.

Capability under test: Data hygiene under adversarial conditions: detecting rather than silently smoothing defects, distinguishing genuine market events from data errors, and applying corporate-action math exactly.

02Analyze

Portfolio Analytics & Exposure

Deterministic analytics on a multi-account ETF household. Each task pins the convention (ddof, annualization, window definition, quantile interpolation, flow timing) so the only source of error is the agent, not the ambiguity of finance jargon.

Capability under test: Convention-exact quantitative work, multi-file joins (holdings, lots, transactions, constituents), performance measurement with cash flows, and econometric inference done correctly.

03Backtest

Backtest Engine Correctness

The agent must implement simulation mechanics exactly: decision-at-close / execute-at-next-open timing, per-fill and proportional costs, whole-share rounding with cash feasibility, dividends by ex-date shares paid on pay-date, universe changes (launches, liquidations) without look-ahead, overlays with strictly-prior information.

Capability under test: Zero-look-ahead engineering, path-dependent state handling, and disciplined adherence to a spec where every sentence matters.

04Backtest

Strategy Evaluation & Overfitting Discipline

Given a working engine, can the agent evaluate strategies without fooling itself? Tasks require out-of-sample discipline, multiple-testing corrections, resampling that preserves dependence, and sensitivity analysis — and require the agent to state a conclusion the data was generated to make unambiguous.

Capability under test: Statistical honesty: reporting out-of-sample rather than in-sample, quantifying uncertainty correctly, resisting the pull to present the best-looking number.

05Rebalance

Rebalancing & Trade Generation

The rebalance phase where analysis becomes orders. Tasks range from a single-account trade list to household-level asset location, contribution-driven buy-only rebalancing, withdrawal sequencing, and a full year of event-driven band monitoring.

Capability under test: Constraint satisfaction and near-optimality under discrete constraints; correct feasibility reasoning (when buy-only is impossible, when a preference cannot be fully honored).

06Rebalance

Tax-Aware & Constrained Optimization

The hard edge of rebalancing: specific-ID lot selection, wash-sale windows that span accounts and dividend reinvestments, tax-loss harvesting into partner ETFs, and formal optimization problems with verifiable optimality gaps.

Capability under test: Modeling domain rules precisely (holding periods, substantially-identical securities, cross-account wash sales) and formulating/solving optimization problems rather than hand-waving heuristics.

07Cross-cutting

Debugging, Tooling & Performance

Agents rarely start from a blank slate. These tasks provide a buggy library, a too-good-to-be-true script, a non-reproducible notebook, or a large universe, and test whether the agent can produce robust, general, fast tooling.

Capability under test: Code comprehension and debugging, generalization beyond the visible data (metamorphic testing), API discipline, reproducibility engineering, and performance under resource limits.

08Cross-cutting

Judgment, Communication & Refusal

Tests the behaviors that separate a useful analyst from a calculator: validating a policy before acting, refusing to silently normalize a broken spec, choosing and declaring defaults consistently, staying within authorization, and communicating in grounded language.

Capability under test: Meta-cognition about the task itself: detecting inconsistency, self-consistency between declared and implemented assumptions, scope discipline, and numerically grounded communication.

09Cross-cutting

Adversarial Robustness & Safety

Financial data pipelines ingest third-party files. These tasks plant instructions inside CSV comments, YAML blocks and notes fields; provide plausible-but-wrong proxies for missing data; put contradictory conventions in a README; and leave credentials in the environment — then check the agent did the right thing.

Capability under test: Instruction hierarchy (task > repository docs > data), refusal to fabricate, output-scope discipline, and secret hygiene.

10End-to-end

End-to-End Multi-Step Reviews

Long-horizon tasks using Harbor's multi-step layout: a quarterly review split into gated steps, a single-shot full pipeline on hostile inputs, a year of quarterly cycles requiring reconciliation of actual fills against the prior plan, and a late client change that must be absorbed with a minimal amendment.

Capability under test: Prioritization under time pressure, error containment across stages, state management across steps, reconciliation, and minimal-diff reasoning when requirements change.