optimalportfolios 5.4.3

Implementation of optimisation analytics for constructing and backtesting optimal portfolios in Python

optimalportfolios — solver hardening drop-in

Adds unconditional post-solve feasibility validation to every optimiser in optimalportfolios, threads the rebalance date through for diagnostics, and removes the deprecated pandas future.no_silent_downcasting option from constraints.py.

Latest additions — pre-solve input contract + observability

This drop also adds (see CHANGELOG_input_contract_and_observability.md for the symbol-level detail):

1. Pre-solve input contract (validate_solver_inputs in solver_diagnostics.py, wired into the TRE wrapper, gated by OptimiserConfig.validate_inputs, default True). Validates the covariance (finite / symmetric / right dimension), flags ill-conditioning (reuses check_covar_conditioning), checks constraint self-consistency (box caps reach full investment, group bounds reachable given the box, benchmark within bounds), and notes dropped assets — at entry, before the solve. An opt-in deep_feasibility=True runs the elastic min-violation LP (diagnose_infeasibility) as a definitive pre-flight. Findings are emitted as one structured InputContractRecord per solve (ERROR for hard covariance problems; conditioning / benchmark / group-reachability folded into a single INFO line) and aggregated by InputContractSummary into one line per category — so a run firing on every rebalance reports "ill-conditioned on N/M; most frequent pair …" rather than N log lines.
2. Relaxation tally — the per-rebalance frozen-overhang relaxation moved off warnings.warn onto the constraints logger at INFO (file, not console spam), carrying a structured RelaxationRecord; a RelaxationSummary handler aggregates them into one run-level line.
3. Relaxation magnitude bound — OptimiserConfig.max_constraint_relaxation (default None); a single relaxation above it escalates the log to ERROR.
4. Environment banner — log_environment() logs python + clarabel / cvxpy / numpy / pandas / scipy versions once per run for reproducibility.
5. RunDiagnostics bundle — configure_run_logging(attach_summary=True) now returns a bundle holding both the rejection and relaxation summaries; .summary() combines them and .check_fallback_gate(...) delegates (existing call sites keep working).

The 20 other warnings.warn sites in constraints.py (rare edge / API-misuse cases, not per-rebalance spam) are intentionally left as warnings; migrating them for channel purity is a separate follow-up.

How to apply

Option A — overlay (simplest). Copy the optimalportfolios/ tree in this archive over your package root; it overwrites the 10 modified files and adds the 3 new ones:

cp -r optimalportfolios/ /path/to/your/repo/

Option B — patch. From your package root (the dir that contains optimalportfolios/optimization/), apply the unified diff. Note solver_diagnostics.py and the two test files are new, so copy those three from this archive first, then:

git apply wiring.diff      # the 10 modified files; review first with: git apply --stat wiring.diff

Reviewing constraints.py: its diff is large (~150 lines) but almost entirely whitespace — removing the with pd.option_context(...) wrapper de-indents its body by one level. Review the real change with git diff -w, which collapses it to the relaxation-logger lines plus the removed with.

File manifest

optimalportfolios/optimization/
  solver_diagnostics.py                         [MOD]  validators + conditioning + infeasibility diagnosis + input contract + run-level summaries/gate/banner
  constraints.py                                [MOD]  relaxation -> structured logger record + magnitude bound; pandas option removed
  config.py                                     [MOD]  diagnose_infeasibility, validate_inputs, max_constraint_relaxation fields
  general/
    quadratic.py                                [MOD]  validate_solution
    max_sharpe.py                               [MOD]  validate_solution (Charnes-Cooper + scipy fallback)
    risk_budgeting.py                           [MOD]  validate_pyrb_solution + validate_scipy_solution + inline guard
    carra_mixture.py                            [MOD]  validate_scipy_solution + inline guard
    max_diversification.py                      [MOD]  validate_scipy_solution
  saa/
    max_return_target_vol.py                    [MOD]  validate_solution x2
    min_variance_target_return.py               [MOD]  validate_solution x2
  taa/
    maximise_alpha_over_tre.py                  [MOD]  validate_solution x2 + pre-solve input contract + relaxation magnitude wiring
    maximise_alpha_with_target_yield.py         [MOD]  validate_solution
  tests/
    solver_diagnostics_test.py                  [MOD]  validators + records + gate + input contract + relaxation tally + banner
    maximise_alpha_over_tre_hardening_test.py   [NEW]  end-to-end integration tests on the wired TRE solver
    frozen_overshoot_relaxation_test.py         [MOD]  relaxation now asserted via caplog + magnitude-bound test

What solver_diagnostics.py provides

• validate_solution(weights, problem_status, constraints, n, *, solver, context, ...) — CVXPY backends. Rejects None / hard-status / non-finite / budget / box-or-long-only violations; accepts optimal_inaccurate iff feasible.
• validate_scipy_solution(weights, res, constraints, n, *, solver, context, ...) — scipy backends. Reads res.success/status/message.
• validate_pyrb_solution(weights, constraints, n, *, solver, context, converged, ...) — pyrb ConstrainedRiskBudgeting (validates .x).
• check_covar_conditioning(pd_covar, ...) — diagnostic only; never modifies Σ.

All three share one feasibility core, so accept/reject logic is identical across backends. On rejection each returns a feasible fallback (drifted weights_0 → benchmark → zeros), logs via logging.getLogger("optimalportfolios.optimization.solver_diagnostics"), and emits warnings.warn on hard rejects. Module dependencies: numpy / pandas only.

Coverage

17/17 solve sites guarded (9 CVXPY + 1 scipy-Sharpe via validate_solution; 3 scipy via validate_scipy_solution; 1 pyrb via validate_pyrb_solution; 2 non-portfolio/array helpers via inline res.success guards).

Tests

python -m pytest optimalportfolios/optimization/tests/ -q
# 36 passed   (also clean under -W error::FutureWarning, confirming the pandas fix)