v0.2.0 - Forecast provider swap and LLM-only robustness

Everything since v0.1.0: the forecast stack now runs on a single free hosted provider, the app is fully functional with just an LLM key, and dependencies are current.

Forecast provider swap

• Bako (TimesFM) is the hosted forecast link. TSFM.ai's API went dark, so the chain now uses Bako's per-symbol ML price predictions (TimesFM on Vertex AI, refreshed daily; free plan 100 calls/day). The run footer labels it bako-timesfm.
• TimeGPT removed. The metered Nixtla integration is gone entirely: no more mid-month quota failures. The fallback is a local log-linear trend fit computed from the same yfinance closes. BAKO_API_KEY is the one forecast key; old saved runs keep rendering their TimeGPT labels.
• Forecast coverage is unchanged: the predicted close lands at the run's outlook horizon (next session, five days, or about a month).

Robust LLM-only mode

• All six agents are populated with only an LLM key. The sentiment agent was the last empty seat: when no social source is configured, it now reads a deterministic crowd proxy computed from the same yfinance data (5-day and 21-day momentum, relative volume, distance from the 6-month high), clearly labeled as a proxy rather than real chatter.
• yfinance covers prices, news, and fundamentals at zero provider keys; the run quality panel and footer name every degradation honestly.
• Backtest replays are untouched: snapshots never carry the proxy, so recorded decisions stay byte-identical.

Dependency and CI updates

• Grouped Dependabot bumps: fastapi 0.156, uvicorn 0.54, crewai 1.15.30, pydantic 2.12.5, python-dotenv 1.2.4, playwright 1.64; actions/checkout v7 and astral-sh/setup-uv v7 in CI.
• The crewai 1.15 line caps pydantic below 2.13, and the lock now resolves that correctly instead of silently downgrading crewai.

Testing

• 33 automated check groups including a real-browser smoke test, all green; CI on Python 3.12 and 3.13.
• New coverage: Bako and crowd-proxy normalizer math, an LLM-only end-to-end scenario, provider-label mapping, and the proxy degradation note in the quality panel.
• Live-verified: full NVDA forecast through the hosted chain.

Paper trading only: educational project, not investment advice.

v0.1.0 - First tagged release

First tagged release of BetterTradingAgents: a FastAPI + CrewAI multi-agent stock research and paper-trading platform.

Highlights

• Six parallel analysts (market regime, technicals, fundamentals, news, sentiment, forecast), bull/bear debate with a neutral judge, and a risk-gated BUY/HOLD/SELL call with sizing and reasoning trail.
• Paper trading through Alpaca (paper API only), with a drawdown brake, position caps, autopilot sessions, and order reconciliation.
• Every decision graded against realized returns and alpha vs SPY; lessons feed later runs.
• Walk-forward backtesting harness, point-in-time data, graded vs SPY after costs.
• Any OpenAI-compatible LLM endpoint, per-role and per-agent overrides, hosted forecasts (TimeGPT / Chronos-2) with local fallback.
• Web UI in vanilla HTML/CSS/JS (no build step), incognito market-only runs, and a full in-app settings page (keys in the OS keychain).
• 33 automated check groups plus a real-browser smoke test; CI on Python 3.12 and 3.13.

Community and security

• Security policy with private vulnerability reporting, issue and PR templates, Dependabot updates (uv + GitHub Actions), and a code of conduct.

Paper trading only: educational project, not investment advice.