qis 5.37.0

qis 5.37.0 adds supplied-covariance Gaussian portfolio risk analytics and includes the JOSS reviewer fixes for missing-price holdings, accounting alignment and paper clarity.

Added

• Add supplied-covariance Gaussian Monte Carlo VaR through GaussianVarSpec,
  GaussianMonteCarlo and compute_gaussian_portfolio_risk. The benchmark-relative
  snapshot reuses RiskModel for volatility, beta and tracking error, with an independent
  closed-form VaR control and explicit horizon units. Existing price-based VaR APIs are unchanged.

• Export the qis.discrete_portfolio subpackage in qis.__all__, so import qis binds it and
  from qis import * includes it. Its names stay qualified as qis.discrete_portfolio.<name>
  rather than wildcarded into the top-level namespace. Existing imports are unchanged.

Fixed

• Apply prior-close weights in date order in to_portfolio_returns on a shared, unique
  DatetimeIndex without NaT, preventing descending or shuffled rows from using a future
  decision for an earlier return. Output row order, caller inputs and missing-term accounting
  are unchanged.

• Reconstruct PortfolioData.instrument_pnl on last observed marks
  (prices.ffill().pct_change()), matching the backtester's convention from #189, so a
  resumed quote books its full move to units held through the gap and instrument
  contributions sum to the nav return. Complete price panels are unchanged.

• Keep plot_bars grouped Seaborn bars tied to their original series colours when an entire
  series is missing, preventing surviving bars from disagreeing with the legend. Bar spacing,
  short-palette cycling and explicit legend overrides are unchanged. MultiIndex columns
  retain their established positional colour mapping after melting to outer-level labels.

• Retain long and short holdings at their last observed valuation mark when execution prices
  are missing in the backtester (#189). Unquoted positions cannot trade, even to a zero target;
  available targets use full marked NAV. NAV, effective weights, carry and instrument P&L now
  follow the same marks, and skipped trades have zero costs. Warn on held stale marks, including
  trailing gaps, and document explicit settlement/exit inputs for terminal events.

• Reject mismatched observation indexes in fit_multivariate_ols before converting inputs
  to positional arrays, preventing fits that silently pair different dates. Aligned fits,
  finite-row filtering and prediction labels are unchanged.

• Align complete, uniquely named instruments_carry panels to price-column order in
  backtest_model_portfolio, preventing one instrument's carry rate from being credited to
  another. Carry timing and existing nonmatching-schema behavior are unchanged.

• Reject conflicting total-group labels in get_group_dict and its grouped-panel and
  portfolio-exposure callers, preventing a single group from silently replacing the
  all-instrument total. Distinct total labels and grouping without a total are unchanged.

Changed

• Reduce peak memory in bootstrap_price_data for float64 list-output draws by replacing
  sampled-return arrays with reconstructed price paths instead of retaining both collections.
  Sampling, price reconstruction, output types and caller-owned inputs are unchanged.
• Reduce peak memory in estimate_rolling_ewma_covar by applying availability masks only
  to returned rebalancing matrices instead of copying the full covariance history. Covariance
  estimates, per-asset warm-up, dates and labels are unchanged; the estimator still retains its
  full covariance tensor.

Documentation

• Clarify the JOSS paper's motivation, software comparisons and research impact, and describe
  currency hedging, factor risk and stress testing in response to reviewer feedback.
• Add a Gaussian VaR monitoring guide and an offline portfolio example.
• Refresh the public API, consumer-usage and paper audit records for the release.

Install from PyPI:

python -m pip install --upgrade qis==5.37.0

The release includes the fix for #189. Runtime dependency requirements are unchanged from 5.36.1.