Changelog

Automatically generates stress scenarios and notifies subscribers when market conditions cross configurable thresholds. Each alert includes an LLM-written cause analysis and a public share link.

Routine security patches and dependency upgrades across the production stack.

Stability improvements across the scenario generation pipeline, intelligence views, and dashboard components.

Reference assets now show 90-day sparklines with labeled 1D/1W/1M performance windows. Macro indicators display native-cadence sparklines.

New dashboard widget showing the probability of entering a stress regime over the next 21 days, conditioned on current macro covariates using time-varying transition probabilities.

Repaired dashboard widgets (media attention, earnings, commodity sparkline), resolved news-pipeline timeouts, and hardened data providers against rate-limit cascades and empty-result caching for more reliable scenario runs.

Intelligence views now include filterable tables for signals and contested claims, making it easier to drill into the evidence behind a scenario.

Scenario intelligence now pulls more relevant insights from SEC filings, with sharper relevance ranking and broader historical coverage.

Secondary-shock correlations are now calibrated per regime using empirical Kendall's tau, with copula parameters surfaced in the copula table.

Scenarios that finish with empty or degraded output are persisted as failed instead of completed.

Long-running scenario pipelines no longer block fast endpoints — they execute on a separate worker process.

Historical scenarios now include a per-factor Market Data view with sparklines, peak/trough readings, and distribution stats for each tracked asset.

SEC filings on the scenario page now show the full detailed summary with a Show more / Show less toggle.

The Predictions tab now loads every attached market via infinite scroll, no longer capped at 30.

Hypothetical scenarios now generate roughly 40% faster end-to-end.

A new dashboard card ranks six stress themes (recession, tariff, inflation, war, uncertainty, cybersecurity) by news-volume z-score versus the trailing 90 days.

Expanded risk factor coverage to include international rates, emerging-market equity, additional commodities, FX, and sovereign bonds.

Shipped 30+ bug fixes from weekly branch-test verification cycles.

Scenarios can now draw on more market-data sources across prediction markets, fundamentals, and macro indicators.

The Statistics tab now shows regime-conditional correlations and surfaces copula family, primary driver, and theta on every secondary shock.

Scenarios now open with a causal narrative (trigger → transmission → end state). New macro widgets added to the dashboard.

Overhauled the scenario page — no more drawers taking the middle of the screen. New Intelligence tab displays extracted risk signals and insights.

Subscription tiers have been replaced with per-scenario credits.

Fixed an off-by-one in historical market data fetching that shifted start dates earlier than requested, distorting factor context, peak-to-trough metrics, rolling shocks, and historical time-series charts. All values now align with the scenario's actual start date.

Bumped cryptography 46→47 and pyarrow 23→24 (majors), plus polars 1.38→1.40 and yfinance 1.1→1.3 for upstream security patches and faster tabular processing.

When the economist falls back to copula-derived shocks, the `fallback_used` flag is now recorded on each shock proposal and surfaced in audit metadata. Removed a redundant `correlation_valid` field that duplicated the master `validation_passed` boolean.

Every context agent (macro, data, filings, prediction, web search) now reports which providers succeeded and which failed. Silent zero-result paths are now visible in audit output.

RSS source is now formally recognized; the news context-agent floor was raised to 120s and partial sub-query results survive timeout. Per-feed failures are surfaced inline (e.g., "rss(2 failed)") instead of disappearing silently.

The Causal Map tab had been silently hidden after a schema rename. It now renders for every scenario that has transmission DAG data.

Every scenario now exposes a transmission DAG mapping how the initial shock propagates through risk factors. Known crises use expert-authored DAGs; novel scenarios get LLM-inferred ones.

A new dashboard card surfaces fresh headlines from 12 curated publishers across political, geopolitical, financial, and policy domains, refreshing every minute. News research in scenario runs is also broader thanks to the new RSS source.

Scenarios can now model recovery regimes — post-crash rebounds, post-election rallies, post-downgrade recoveries — with sign-agnostic shocks. The historical knowledge base gained 6 recovery narratives spanning 1987 to 2020.

Historical scenarios now dynamically retrieve the most relevant precedents per run from an expanded knowledge base, replacing a hardcoded list. The taxonomy gained 9 recovery archetypes and 2 new crisis archetypes (banking crisis, asset bubble).

Filings agent now extracts per-filing risk factors from Item 1A; macro agent grounds historical comparisons with statistical anomaly detection; news agent surfaces its sub-query plan; data agent uses deterministic per-asset volatility regimes.

Bounded per-provider news timeouts prevent a slow data source from blocking runs; the economic calendar widget no longer errors on an invalid parameter; and the pipeline now recovers correctly when an intermediate LLM call returns 0 shocks.

Agent prompts are now deduplicated via shared Jinja2 fragments, a redundant context-synthesis LLM call was removed, and Kalshi + Polymarket agents share a common prediction base class. Scenarios generate with fewer round-trips and more consistent framing.

A new context coordinator dispatches only the agents (news, macro, filings, prediction markets, etc.) most relevant to each scenario, with per-agent query hints. Result: sharper, less noisy context and better evidence selection on every run.

A new LLM-as-judge step cross-checks proposed shocks against historical analogs, percentile anchors, and copula-implied bounds — producing a `judge_plan` with hard `[min, max]` constraints before the scenario reaches the quantitative layer. Surfaced in the API response and scenario detail UI.

Ten new dashboard cards — Correlation Heatmap, Rate Expectations, News Sentiment, Scenario Timeline, Watchlist, Portfolio P&L, MOVE Volatility, Global Macro Calendar, Scenario Comparison, and Provider Health — give you a richer at-a-glance view of market and pipeline state. Each card includes an empty state, staleness indicator, and refresh timestamp.

Kalshi and Polymarket are now analyzed by dedicated agents with LLM-powered market interpretation, replacing the single prediction agent for more granular signals.

News agent now plans and executes multiple search queries in parallel, then aggregates results for faster and more comprehensive coverage.

US Treasury 2Y yield now correctly sources from FRED instead of Yahoo Finance futures.

Added EWMA correlation estimation, Ledoit-Wolf matrix shrinkage, and condition number monitoring for more robust risk factor dependency modeling.

New dashboard widget tracking NFCI, STLFSI4, and OFR FSI financial stress indices.

Users can now share any scenario to a public page.

Key metrics in historical narratives now show provenance tooltips revealing the exact data source and computation method behind each number.

When a crisis isn't in the knowledge base, an agentic research loop (news/web search) fills the gap automatically.

New contact form page for reaching the StressGen team. Added a curated regulatory calendar with verified dates for upcoming Fed, ECB, and prudential deadlines. Crisis Chronicle drawer now surfaces historical narrative context.

Fixed HMM regime detector to correctly identify market stress states in the dashboard. Improved site discoverability with SEO fixes across sitemap, metadata, and llms.txt. Dashboard widget preferences now persist correctly across sessions.

Added GDELT and Finnhub as new news providers for broader market and geopolitical coverage. Prediction market data from Kalshi and Polymarket now injects price momentum signals into scenario generation with tighter relevance filtering.