Run and deploy your AI Studio app

This contains everything you need to run your app locally.

View your app in AI Studio: https://ai.studio/apps/d6dd78f9-8412-4bd7-bf1b-76dad78f86de

Run Locally

Prerequisites: Node.js

1. Install dependencies: yarn install
2. Create an ignored .env.local file for local secrets. Set GEMINI_API_KEY there, and use CI/deployment secret storage for hosted environments.
3. Refresh the cached market data: yarn run update:market-data
4. Run the S&P 500 statistical relevance gate: yarn run backtest:market
5. Run the app: yarn run dev

Programmatic API

Run the local Express API:

yarn api

Endpoints:

• GET /health — liveness check.
• GET /api/assets — supported forecast assets.
• GET /api/forecast?asset=btc&horizon=180&confidence=0.95 — compact forecast summary for btc, sp500, or gold.

The API is implemented with lightweight Express decorators in src/server/decorators.ts; controllers use @Controller and @Get.

Market Tabs

The app defaults to the BTC tab and preserves the Bitcoin power-law forecast workspace. The S&P 500 tab uses VOO as an investable S&P 500 proxy because it has daily OHLCV and volume data; VTI is deferred as a broader-market option.

VOO data lives in src/data/voo-history.json and is regenerated with npm run update:market-data. The updater uses Yahoo Finance's no-key chart endpoint and normalizes OHLC values with the adjusted-close ratio when the source provides adjusted close.

The S&P 500 forecast uses a statistical log-return model:

equity_premium = clamp(expanding mean daily log return, 0.00005, 0.00055)
drift = equity_premium - 0.25 * mean(r_90) + 0.25 * mean(r_252) + 0.10 * trend_252
volatility = 0.65 * stdev(r_90) + 0.35 * stdev(r_252)
median_h = close * exp(drift * h)
interval_h = median_h * exp(+/- z * volatility * sqrt(h))

The S&P 500 lower/top lines reuse the same chart fields as BTC floor/peak lines, but use a VOO-specific statistical channel instead of Bitcoin power-law curves:

trend = 126-session SMA
residual = log(close / trend)
lower = trend * exp(2.5th percentile residual over prior 1,260 sessions)
top = trend * exp(99th percentile residual over prior 1,260 sessions)

npm run backtest:market is the reproducible statistical gate for the S&P 500 model. On the current VOO cache through 2026-06-05, the walk-forward channel test covers 96.3% of sampled closes with 2.4% below-channel breaks and 1.4% above-channel breaks. The median forecast also passes at 30, 90, and 180 trading-day horizons with statistically significant median-error improvement versus a no-change baseline and directional relevance against a 50% null.

Daily production market quotes

BTC, VOO, and GLD retain checked-in JSON as an immediate read fallback. Production daily candles are refreshed by workers/market-quote-refresh at 23:15 UTC and stored in D1. Cron changes can take several minutes to propagate.

Create three separate databases (local Wrangler state, preview, and production), then replace the placeholder IDs in both Wrangler configs. Never point tests or preview at the production ID. Deploy in this order:

1. npm run market-quotes:migrate:preview
2. Deploy the Worker without enabling its cron and invoke a local/preview scheduled event twice; inspect market_candles and refresh_runs for idempotency.
3. Deploy Pages with the D1 binding and verify /api/market-data and /api/forecast return the same latest candle.
4. npm run market-quotes:deploy:preview, then enable the production cron only after the preview proof passes.

Use wrangler tail --config workers/market-quote-refresh/wrangler.toml for Worker logs and query refresh_runs for per-asset outcomes. Run MARKET_DATA_BASE_URL=https://preview.example npm run market-quotes:smoke for the cross-endpoint freshness check. The independent GitHub watchdog is enabled only after setting the repository MARKET_DATA_BASE_URL variable.

Rollback is non-destructive: disable the Worker cron and remove/disable the Pages D1 binding/read path. The application then reports fallback and continues from bundled JSON. Do not delete D1 rows during rollback.