BTC–Jev Signal

A public, non-trading BTC forecasting experiment using TypeSafe Jev with public Kraken spot and Binance futures market data.

Public dashboard: https://lab.rokogrga.com/btc-jev

Cloudflare fallback: https://btc-jev-signal.roko-experiments.workers.dev

This repository owns the complete BTC–Jev project: its dashboard, market-data collection, Jev judgments, scoring, Cloudflare Worker, and D1 schema. The separate WebGrga/rg-lab repository owns the public project index and routes /btc-jev to this project's independent Netlify deployment.

This is a personal software experiment, not financial advice, investment research, or a trading service. It never places or prepares trades.

What the experiment does

At exact UTC boundaries, ordinary code collects a neutral structured market State. Jev answers atomic higher-or-lower questions with full probability distributions. Every forecast is frozen with its issue price and exact target timestamp. After the target passes, ordinary code retrieves the completed 1-minute close from the same spot source and calculates the outcome.

The primary experiment uses a natural, non-overlapping schedule:

Forecast	Issued	Target	Maximum primary trials per UTC day
15 minutes	Every 15 minutes	Issue + 15 minutes	96
1 hour	At the top of each hour	Issue + 1 hour	24
4 hours	Every fourth UTC hour	Issue + 4 hours	6
UTC day close	At 00:00 UTC	Next 00:00 UTC	1

An earlier version issued every horizon every 15 minutes. Those append-only records remain auditable, but overlapping 1-hour, 4-hour, and day-close forecasts are excluded from the primary dashboard scores unless their issue time matches the natural schedule.

What Jev receives

Every State uses normalized UTC ISO 8601 timestamps and completed candles only.

Measurement	Source and definition
Anchor and realized target	Cloudflare: Kraken Spot completed BTC/USD 1-minute candle. Local runner: Binance Spot completed BTCUSDT 1-minute candle.
Returns	1m, 5m, 15m, 1h, and 4h from the runtime's spot source
15m / 1h / 4h indicators	RSI(14), MACD(12,26,9), ATR(14), SMA/EMA 20/50/200, price distances, bar returns, and volume ratios from completed spot candles
UTC session	Spot open, high, low, volume, return from open, and minutes to 00:00 UTC
Perpetual futures	Funding, mark/index price, open interest, and OI changes from Binance USD-M Futures
Liquidations	Short timestamped observation of Binance's public BTCUSDT force-order WebSocket
Order book	Cloudflare: Kraken Spot top-20 levels. Local runner: Binance Spot top-20 levels. Both expose bid/ask notionals, spread, and imbalance.

The State always identifies snapshot.symbol, snapshot.quote_asset, and every source string. Some stable JSON field names still end in _usdt for backward compatibility; in Cloudflare records those monetary values are USD, as declared by quote_asset: "USD".

The 15m, 1h, and 4h questions share one State and run independently in one TypeSafe request when they are due together. The day-close question is a second-stage request that can consume the three horizon distributions. The API key stays server-side.

Reading the dashboard

The dashboard intentionally separates four concepts:

1. Active forecasts show the newest eligible prediction for each horizon, its reference price, target, probability split, and Jev confidence.
2. Forecast lifecycle shows exactly how observed data becomes calculated State, a frozen prediction, and a scored outcome.
3. Performance by horizon keeps different time windows separate. Accuracy measures the selected direction. Brier score and log loss evaluate probability quality; lower is better.
4. Exact inputs exposes every major field and includes the complete JSON State for audit.

Small samples are descriptive only. Jev confidence describes concentration in the returned distribution, not guaranteed correctness.

Cloudflare production architecture

• Cloudflare Workers serves the API and runs the scheduled experiment.
• A project-specific Netlify site builds this repository's dashboard under /btc-jev; the RG Lab edge route exposes it at the canonical lab.rokogrga.com/btc-jev address.
• Cloudflare Static Assets also provides a fallback copy on workers.dev.
• Cloudflare D1 stores prediction batches and settlements.
• Cloudflare Worker Secrets stores TYPESAFE_API_KEY.
• One Cron Trigger runs at minutes 1, 16, 31, and 46, allowing the just-completed Kraken candle to finalize before collection.

The production database is intentionally separate from data/*.jsonl. A new deployment starts a clean online history unless a local history import is deliberately approved and performed.

Deploy

Requires Node.js 20.19 or newer and a Cloudflare account.

npm install
npx wrangler login
npm run cloudflare:migrate
npx wrangler secret put TYPESAFE_API_KEY
npm run cloudflare:deploy

Cloudflare configuration is in wrangler.jsonc; the D1 schema is in migrations/.

Useful Cloudflare commands:

npm run cloudflare:check   # Build and validate without publishing
npm run cloudflare:dev     # Local Worker/D1 development
npm run cloudflare:deploy  # Build and publish

Local experiment

Copy .env.example to .env, add TYPESAFE_API_KEY, and run:

npm install
npm run experiment

The local runner uses the same natural cadence and writes append-only records to data/. Restarting is safe; existing boundary IDs are not duplicated and overdue forecasts are settled on startup.

Other useful commands:

npm run dashboard         # Read-only dashboard at http://127.0.0.1:3000
npm run experiment:once   # Immediate four-horizon diagnostic batch
npm run experiment:state  # Inspect State without calling Jev
npm run settle            # Settle due local forecasts
npm run report            # Regenerate local report files
npm run check
npm test

Stored data

Local mode writes:

• data/predictions.jsonl: complete States and probability distributions.
• data/settlements.jsonl: target prices, realized returns, accuracy, Brier score, and log loss.
• data/report.json and data/report.md: aggregate summaries.

Cloudflare stores equivalent JSON records in indexed D1 tables. Credentials are never written to State, dashboard responses, or prediction records.

Reliability

• Market HTTP requests use bounded retries and timeouts.
• TypeSafe requests use SDK retries and respect rate-limit guidance.
• Optional derivatives, order-book, or liquidation failures remain explicit in State.
• Forecast and settlement IDs are unique, making repeated scheduled delivery safe.
• The public API is read-only and returns no secret or TypeSafe usage metadata.
• Dashboard and API responses include restrictive security headers.

Netlify dashboard deployment

This repository deploys only the BTC–Jev dashboard. netlify.toml builds web/dist, serves the app beneath /btc-jev, proxies /btc-jev/api/* to the Cloudflare Worker, and redirects the project deployment root to the canonical RG Lab address. The hub and its cross-project routing live in WebGrga/rg-lab.