v0.7.2

0.7.2

Changed

• fit_negbinomial now refuses non-overdispersed data. The negative
  binomial always satisfies variance/mean = 1/p > 1, so when the sample
  variance is not greater than the mean no finite (r, p) maximizes the
  likelihood -- the supremum is the Poisson limit (r -> inf, p -> 1). The
  fitter previously ran the optimizer anyway and returned an arbitrary,
  non-optimal finite result (with success=True) whose value was unstable
  across versions and platforms. It now raises ValueError pointing to
  fit_poisson, matching the long-standing contract of
  fit_negbinomial_moments. Because fit_best_frequency already drops
  candidates that fail to fit, model selection on underdispersed data now
  cleanly falls back to Poisson.

Removed

• Obsolete negbinomial fit warning suppression. The np.errstate guard
  added in 0.7.1 silenced a SciPy finite-difference RuntimeWarning that only
  arose when the optimizer chased the p -> 1 boundary on near-Poisson data.
  With such data now rejected up front, the optimizer only runs on
  well-conditioned interior problems and the warning no longer occurs, so the
  suppression has been removed rather than carried forward -- the guard is the
  root-cause fix.

Tests

• Extended tests/estimation/test_negbinomial_parameterization.py to pin the
  new refusal (underdispersed, equidispersed, and degenerate inputs raise),
  confirm overdispersed fits are warning-free without suppression, and verify
  fit_best_frequency falls back to Poisson on underdispersed data.

v0.7.1

0.7.1

Changed

• fit_negbinomial performance. The Negative Binomial MLE now aggregates
  the data to distinct counts and scores each optimizer step with a single
  vectorized scipy.stats.nbinom.logpmf call weighted by the count
  multiplicities, instead of evaluating a scalar PMF once per observation per
  iteration. Fitted results are unchanged (the likelihood is identical up to
  floating point); only the runtime differs -- the two official
  fit_best_frequency model-selection tests drop from ~23.5s each to well
  under a second. Scoring in log space via logpmf also avoids the tail
  underflow of the previous log(pmf(...)).
• fit_negbinomial warning hygiene. The L-BFGS-B finite-difference
  gradient probes the edge of the feasible (r, p) box, where the negative
  log-likelihood is +inf; the resulting inf - inf inside SciPy's numerical
  derivative raised a benign numpy "invalid value encountered" RuntimeWarning
  during otherwise-successful fits. That specific floating-point warning is now
  suppressed for the duration of the optimization only.

Fixed

• examples/panjer_vs_simulation.py aggregate grid. The Panjer recursion
  step count is now sized to the discretized severity support
  (int(round(max_loss / h))) rather than a fixed 50,000. The old value
  computed the aggregate PMF out to ~250 aggregate means at large cost with no
  effect on the reported mean; the example now runs an order of magnitude
  faster while reproducing the same result.

Tests

• Added tests/estimation/test_negbinomial_parameterization.py, pinning that
  NegativeBinomial(r, p).pmf matches scipy.stats.nbinom(r, p) (and not the
  flipped (r, 1 - p) convention), that the fit's logpmf likelihood agrees
  with the model's own PMF, and that the fit stays frequency-aggregated,
  fast, and free of the finite-difference warning.