Model Changelog

Unit-aware PDF parser, scheduler that polls the SharePoint feed every 30 minutes and parses pending PDFs every 4 hours, dual-write semantics that land structured financials in fundamentals and earnings_actuals on every successful parse. Default-to-fail unit detection (missing beats wrong-by-1000×) with quarantine routing. Connection- ordering regression in the overlay pipeline corrected so the v9.2.27 calibrated risk pills (Garman-Klass volatility, distribution- day count, max() composite) re-attach to live rankings.

Architectural pivot from hardcoded MARKET_EXPECTATIONSto database-canonical earnings_actuals. The intelligence-feed parser already ingested releases on a 2-hour cadence; the overlay now reads them. New/api/earnings/freshness endpoint exposes coverage, capture latency, and source-mix metrics with a four-tier health classification monitored by the Data Sentinel agent.

Per-series freshness thresholds widened to match actual NGX publication cadence (CPI monthly with ~15d publication lag; MPC cycle ~85d). The prior universal 1/3/7-day ladder was flagging legitimately-current data as stale on quarterly inputs. Health calculation for the Agent Operations dashboard rate-based with a new WARNING tier between HEALTHY and DEGRADED so transient failures are visible without dominating headline state.

Single risk pill combining 20-day Garman-Klass volatility and 25-day distribution-day count via the max() operator. Five composition schemes tested; only max() cleared the ship threshold. Spearman IC +0.5053 against forward 10-day drawdown, top-vs- bottom quintile separation 5.84×. Documents the structural insight that weighted-average composites dilute strong frontier- market signals; max() preserves the loudest scream without watering down the strong signal toward the weaker one's noise.

Factor stack expanded from seven to nine factors. Weights rescaled uniformly by 0.9× to allocate 10pp to two alternative-data signals: a 120-phrase Nigeria-tuned MD&A lexicon at 8% weight and an earnings-surprise factor at 2% weight. NLP scores are section-weighted (outlook 1.2×, auditor 1.5×, chairman 0.6×) with a negation window that flips polarity on negated phrases.

Regime-conditional ERP premium and margin-of-safety bands. Banking-sector valuation uses a two-stage Damodaran dividend discount model (five-year explicit stage at sustainable growth, Gordon tail at 4% Nigerian long-run nominal GDP floor) because single-stage Gordon goes pathological at the ROE × retention levels Nigerian banks routinely produce. YoY EPS trajectory and sector-aware leverage adjustments catch value-trap configurations that the underlying value composite alone misses.

New sector_map_history audit table seeded from the minimum factor-scores run-date per ticker. The portfolio backtest now queries this table for the universe-as-it-existed at each historical rebalance date, replacing the previous current-snapshot assumption. Survivorship bias eliminated; backtest results now reflect what an operator could have actually traded at each historical point.

Replaced silent 0.5 substitution for missing NGX market-breadth observations with inverse-distance-weighted imputation from nearest months with real data. Per-month breadth coverage tracking; confidence haircut up to 20% when imputed-month share exceeds 30%; regime_uncertainty_flag fires when coverage drops below 50%. The HMM no longer learns corrupted transitions from fabricated neutral observations.

Three of five Port Activity Index components migrated from synthetic/demo data to real feeds: AIS vessel tracking via aisstream.io WebSocket (35% weight), RSS-based news sentiment from Ships & Ports + Nairametrics (8% weight), and freight pressure via FRED Brent crude (25% weight). Dynamic confidence ceiling now responds to per-source freshness rather than a hardcoded 5.5 / 10 cap.

Three-state breaker (CLOSED / OPEN / HALF_OPEN) on the factor- response builder with last-known-good persistence and a four-tier degradation chain (fresh → cached → raw-DB → 503-recovery). Five downstream consumers (/api/factors, /api/session, /api/retail, /api/rebalance/targets, /api/chat) all serve last-known-good rather than empty when the builder is unavailable.

Hardcoded macro fallbacks replaced with a self-healing loop: scrapers write to macro_data on success; get_seed_value() reads the latest verified row before falling back to a hardcoded last-resort constant. The fallback is now “last successful scraper run” not “last code deploy”, so seeds stay current without operator intervention. New /api/seeds/statusendpoint surfaces per-seed freshness ratings.

Seven fixes targeting model-assessment gaps: MIE insider-dealings staleness recalibrated to the NGX weekly filing cadence; dynamic carry threshold wired to the /api/factors pipeline (previously defined but never invoked); aggregate overlay cap enforcement post-all- overlays; factor IC tracker bootstrap from up to twelve months of historical scores (no synthetic data); weekly cross-source price validation against Yahoo Finance; PAI confidence ceiling computed dynamically from per-source freshness.

Eight-capability ops module: multi-channel alert dispatcher with throttling, data anomaly detector with price-and-macro sanity bounds, circuit-breaker pattern for the database / EODHD / macro APIs, stale-on-error cache, corporate-action detector (split and reverse-split detection from price ratios), enhanced audit wrapper, API response-timing middleware, EODHD price guard. All configurable via Railway environment variables for production alert routing.

Three new overlays (FX direction, regime-conditional signal thresholds, carry-adjusted alpha) close the alpha pipeline to its current six-layer shape. New rebalancing module with three- tier policy: quarterly full rebalance (30% turnover cap), monthly tactical (10% turnover cap with momentum delta), and 24-to-48-hour event-triggered rebalances on CBN MPC, regime flips, GMF, Brent ± 15%, or sovereign rating actions.