WeatherEdge

WeatherEdge forecasts station daily-high temperatures, converts distributions into prediction-market bracket probabilities, and tests fee-aware trading policies in separate paper accounts. It combines weather ingestion, probabilistic calibration, orderbook evidence, simulated execution, official settlement, AWS operations, and a React evidence dashboard.

Live dashboard · Strategy Lab · Repository map · Architecture · Project case study · AI-assisted development

Paper trading only. The engine reads public prediction-market prices and records simulated orders. live_execution.py rejects non-dry orders; no authenticated real-money execution client is implemented. Paper fills do not establish that an equivalent real order would fill.

Current evidence and paper accounts

Repository source covers twenty city markets. V7, M1 and the November 7 capital changes were deployed on November 7, 2026: backend at c14d1a10c, web at 372212b33. The October 11, 2026, 11:01 PDT backend snapshot of 2a6432e3 and fifteen markets is historical. Repository source, installed backend, SPA assets, and generated runtime data have separate provenance; a merged fix is not proof of a deployed fix.

Area	Source capability and observed deployment
Coverage	20 city markets, with SFO as the flagship
Policy	Live, V7 and M1 in separate $2,000 paper accounts; live execution disabled
Website	V7 history/audit and hybrid research UI; November 7 release verified the installed app, Pages deployment and public hashes; locked MapLibre 6.4.1
Backend cutover	Done November 7: V7 and M1 active; V6 archived with monitoring and settlement preserved

At the older October 5, 19:52 PDT account snapshot, the separate $1,000 initial-capital accounts reported:

Paper account	Realized P&L	Return on initial capital
Live Stability	+$58.67	+5.867%
Research ROI v6	+$17.93	+1.793%

Research v6 earned $17.93 on $3,927.86 of resolved entry capital: 0.4564% resolved-capital ROI. Its approximately 72% profitable-exit rate did not imply strong net profit. These accounts are economically separate; their profits and balances must not be treated as one bankroll.

The v7 audit records the measured baseline, release gap, thin fill capture, concentrated losses, stale analysis, and reproduced forecast/exit evidence defects. Source corrections require verified deployment and prospective validation. The November 7, 2026 owner decision sets each active paper account's capital at $2,000: Live Stability, Research V7, and the M1 trial. Live kept its ledger and order history with one audited +$1,000 deposit; V7 and M1 got fresh ledgers. V7's reference equity is $2,000, with no reserve. These accounts are never summed into one bankroll. Archived V6 and older accounts keep their real $1,000 history and are never restated. See the paper limits and completed cutover. Strategy Lab shows the policy version in its artifact and all published archived profiles, including zero-trade experiments.

The fresh forecast baseline records 125,170 NWP member rows and 10,642 source-separated scored cases. Its 1.80°F MAE and 1.30°F CRPS describe reconstructed archive skill; original decision-time availability remains unestablished. The dated incident reports, unassigned losses, and older ML challenger are retained alongside the audit.

The current objective is to gather meaningful evidence for or against profitability. The $2,000 paper capital and V7's 5% ($100) daily goal are research settings, not promised or achieved returns. Older targets remain dated research assumptions. October 31, 2026 was the engineering and evidence review date; readiness depends on the evidence. Increasing capital, requested size, or activity does not create an edge or displayed liquidity. Scaling requires account-scoped net P&L after fees, calibrated probabilities, realizable execution, drawdown evidence, and chronological holdout validation. Live allocation remains $0 until the owner explicitly authorizes real money.

The bounded Mac worker supports separate prospective weather research. As of October 10, its store held 489 immutable forecast distributions; 139 unique station/lead/date outcomes covered five completed dates, including day-ahead diagnostics. This is not net trading profit or a promoted model. The five-city reconstructed training-history bundle was deployed November 7. Two later process-signaling failures remain under investigation. See the October 11 consistency audit and session memory for the dated operational limits and remaining recovery work.

Forecasting and trading path

NWP history + live vintages ► per-station EMOS ──► forecast distribution
Final station CLI truth ────► training and independent scoring        │
Fresh station observations ─► same-day feasible outcomes             ▼
Real orderbook/tape ────────► market posterior + exact fees + risk gates
                                                                  │
                                                                  ▼
                       separate paper accounts ──► AWS JSON ──► React SPA

The shared operational weather path uses eight NWP members and per-station EMOS post-processing. Each city has a defined market series, settlement station, NWS Climatological Report, and fixed-standard climate-day clock. SFO also has legacy blend adapters, station history, LSTM/XGBoost research, and marine-layer features. Those historical studies are not proof of current twenty-city serving accuracy. An EMOS point must be paired with its matching EMOS distribution.

Google Weather runs through a budgeted, provider-expiring private runtime store. The scheduled orchestrator serves the non-Google baseline first. Research challengers and legacy SFO interfaces have explicit boundaries; optional source availability does not establish a promoted forecast improvement. Apple WeatherKit, when enabled, nudges next-day and later forecasts with a fixed, capped weight that is recorded for scoring; see its boundary. Neither provider's historical conditions should be treated as an original decision-time forecast vintage.

Scheduled entry uses portfolio-scan: Live Stability and V7 reservation-price maker limits and configured bounded taker crosses must pass exact fee, edge, lower-bound, liquidity, concentration, loss, and account admission checks; M1 uses its frozen market-recalibration entry rule with separate risk checks. Monitoring and settlement preserve each account's identity. Final settlement uses the city's official final CLI truth; a partial or preliminary report cannot fabricate a resolved outcome.

Evidence library

Evidence	Scope and limitation
v7 audit	October baseline and corrections; prospective v7 profit is unproven
CRPS pilot	2,658 reconstructed fixed-lead forecasts; exploratory, not execution replay
ML challenger	MAE improved 2.34% on the exploratory sample, interval included no gain, CRPS worsened; not promoted
Historical SFO evaluation	Older station/model-specific experiment; not current account ROI or universal forecast skill
Rejected retune	Apparent improvement rejected as statistically weak

Use proper scores such as CRPS, Brier/log loss and calibration alongside point error. Evaluate independent calendar-date clusters and account/policy eras; repeated scans and correlated city outcomes are not independent observations. Readiness evidence belongs to Live Stability, while research histories remain separate.

Local setup and verification

Use Python 3.12 or 3.13 (the project minimum is 3.11):

python3.13 -m venv .venv-dev
.venv-dev/bin/python -m pip install -e '.[dev]'
bash scripts/run_tests.sh
bash scripts/verify_project.sh

The full Python gate runs the repository health check, Semgrep, the backend and forecaster suites, and compilation. Local Semgrep absence produces a warning; CI requires the scan. These checks do not build or validate the web app.

Analysis commands that place no orders, from the repository root:

.venv-dev/bin/python -m sfo_kalshi_quant.cli --no-color analyze --target-date rolling --side both --cities all
.venv-dev/bin/python -m sfo_kalshi_quant.cli --no-color paper-report
.venv-dev/bin/python -m sfo_kalshi_quant.cli --no-color backtest-signals --sample-mode entry-per-market-side

The root pyproject.toml is the sole Python install manifest. It owns the sfo_kalshi_quant package and sfo-kalshi console entrypoint. Heavy model training uses .[train] locally; production uses lightweight dependencies. See forecaster instructions and trading instructions for data and command boundaries.

Web app

The public site is the root React + TypeScript + Vite + HeroUI Pro SPA (src/). Licensed components require HEROUI_KEY and the version-preserving installer used in CI; plain package installation supplies stubs and may not reproduce the app.

bun install --frozen-lockfile --ignore-scripts
env -u CI npx -y [email protected] --auto
git diff --exit-code -- package.json bun.lock
bun install --frozen-lockfile
bun run lint
bun run test
bun run build
bun run preview --host 127.0.0.1 --port 4173

Check desktop and mobile behavior in a browser against that dist/ build. Initial browser-observed resource capture and budgets are separate verification gates:

bun run bundle:report
bun run bundle:capture -- /tmp/weatheredge-initial-resources.txt
bun run bundle:check:observed -- /tmp/weatheredge-initial-resources.txt

The SPA fetches runtime JSON at load. AWS publishes prebuilt assets plus fresh trading_signal.json, forecast_data.json, weather_story_data.json, strategy_research.json, and cities_data.json to GitHub Pages. Freshness and hashes come from publication_manifest.json; HTTP 200 alone proves no data freshness.

Repository and runtime authority

Source belongs to Git; production weather/paper databases and current artifacts belong to AWS. Ignored MacBook state can be stale. Before local dashboard design checks, clear disposable local runtime state with the canonical helper:

python3 scripts/clear_local_runtime_state.py --confirm

This is a local cleanup, not a production data migration. Keep raw data, credentials, private exports, keys, and operator details out of Git. Preserve attribution and dated research records. The repository map identifies canonical entrypoints, intentional mirrored files, and compatibility wrappers. Use the deployment runbook for clean-revision deployment, verified backups, account reconciliation, timer recovery, and publication verification.

MIT — see LICENSE and third-party notices.

Local compute and cost

V7 supports a two-hour Mac research worker for fresh read-only paper/forecast audits and calibration comparisons, with preserved results and job-scoped caffeinate. AWS continues collection and paper execution. Full local replicas use transaction-aware SQLite copying; they do not bypass the durable deployment backup gate. See local compute operations for setup, verification and the distinction between accrued bills and monthly cost.