riskprism 0.9.1

Patch release: version is now derived from package metadata (the 0.9.0 wheel reported a stale 0.1.0). Model artifacts live under model-* releases.

riskprism 0.9.0

Package release. Model artifacts live under model-* releases.

Model build model-2026-08-24

Weekly PRISM-US-MH artifacts. See docs/METHODOLOGY.md. Not investment advice.

Model build 2026-08-22b (PRISM-US-MH-0.9 + SH)

v0.9: growth style, leverage composite, FF30 industries (K=40), coverage 2,987 → 6,307 (estimation pinned at the liquid ~2,800). Mean daily R² ~0.21; scoreboard at ideal (bias 0.99, |z|>1.96 5.2%, 135 weeks). Dividend yield measured and rejected; two failed coverage attempts documented. Short-horizon variant included. Evidence: docs/DECISIONS.md §13–15. Not investment advice.

Model build model-2026-08-22 (PRISM-US-MH-0.7)

v0.7: Market Sensitivity (beta) split from Residual Volatility. Beta significant in 83.7% of daily cross-sections; mean daily R² 0.159 → 0.178; overall bias 1.05 → 1.01. Evidence: docs/DECISIONS.md §12. See docs/METHODOLOGY.md. Not investment advice.

PRISM-US-MH-0.6: value & quality composites

Value and quality rebuilt as multi-descriptor composites (value: book/earnings/cash-flow/sales yields; quality: ROE, ROA, OCF/assets, gross margin), the USE4/Axioma construction. Measured on the same 149 weeks: quality went from significant in 1.6% of daily cross-sections to 49.2%, value 4.1% to 11.7%, mean daily R² 0.156 → 0.159, with the validation scoreboard unchanged (overall bias 1.05, min-var optimized 1.07, MZ slope 1.10). Exposure definitions changed, so history rebuilt cold — 713 daily regressions, 121 scored validation weeks. Full evidence in docs/DECISIONS.md §11.

Artifacts: exposures, factor covariance, specific risk, daily factor returns/residuals/t-stats, validation history, and the distilled point-in-time EDGAR fundamentals store (now including OCF, revenues, gross profit, cost of revenue).

PRISM-US-MH-0.5: daily estimation

PRISM-US-MH-0.5 — estimation switches from weekly to daily returns (weekly formation, daily regressions): ~10× the covariance effective observations (252d correlation half-life ≈ 730 obs). Measured: Mincer–Zarnowitz slope 0.81→1.02, min-var optimized-portfolio bias 1.36→1.09 (matching Shepard's 1/(1−K/N_eff)), ETF bias statistics 0.87–1.08. Also ships: optimized=true Shepard correction flag (Python + MCP), per-factor WLS t-statistics (factor_tstats.parquet + /model.md quality table), daily ETF RBSA. Cold rebuild (weekly history incommensurable). 2,987 covered / 2,774 estimation / 713 daily regressions over 149 weeks.

Provisional demo build (300 names)

PRISM-US-MH-0.2 · 300 covered / 294 estimation · as of 2026-08-14.

Provisional release: the model machinery is final but this build covers only the ~300 largest US stocks. The first full (~3,000-name) build will supersede it. Contents: exposures, factor covariance, specific risk, factor returns, residuals, asset_meta (estimation quality), and the distilled EDGAR fundamentals store (public domain). Not investment advice.