ECHO

ECHO is an event-conditioned market intelligence terminal. It maps real-world events through explicit economic relationships, estimates abnormal-return distributions, preserves every forecast vintage, and acts only through a deterministic risk gate.

ECHO does not predict whether news is bullish or bearish. It estimates how economically meaningful information should propagate through connected companies, measures how much of that impact the market has already absorbed, and acts only on the remaining propagation gap.

Product surface

The UI has five primary destinations: Dashboard, Events, Graph, Forecasts, and Strategy. Backtesting, simulation, ledger inspection, risk, and paper status are subviews rather than additional navigation tabs.

Architecture:

World Monitor ──┐
                ├──> Canonical events ──> ShockGraph ──> Forecast
Market data ────┘                                      │
                                                       ▼
                                                  Risk gate
                                                ┌──────┴──────┐
                                                ▼             ▼
                                             Abstain      Paper trade
                                                              │
                                                              ▼
                                                        Alpaca PAPER

Quick start

Requirements: Docker Compose, Python 3.12+, uv, Node 22+, and pnpm.

make setup
make demo-reset   # clean local PostgreSQL, migrations, seed, graph snapshot
make demo         # deterministic replay UI/API; no live feed or broker calls

Open http://localhost:3000. The global header always exposes data mode (LIVE, REPLAY, MOCK, or OFFLINE) and execution mode (SIMULATION or PAPER). Replay is never labelled live.

For normal development:

make dev          # database, migrations, mock event, API, web
make dev-replay   # same stack with deterministic market replay

Research and safety

The locked resource and technology boundary is documented in docs/resource-plan.md. Optional models and future infrastructure are explicitly separated from the replay-safe core.

• echo-rp-v1 is the frozen strategy artifact; its config hash is recorded in artifacts/release/manifest.json.
• Forecasts are abnormal returns with q10/q50/q90, horizons, cutoffs, graph snapshot identity, feature hashes, and content hashes.
• Ledger and graph snapshot records are append-only and protected by database constraints/triggers.
• Execution defaults to DISABLED; LIVE is rejected. Only validated PAPER_TRADE decisions can reach the Alpaca paper adapter.
• Alpaca paper trading is simulated broker execution. It is not proof of live profitability and does not model all impact, latency, queue, or slippage effects.

Data modes

Set ECHO_MODE explicitly to offline, replay, or live. Market providers use ECHO_MARKET_PROVIDER (off, replay, massive, or alpaca) and world feeds use ECHO_WORLD_PROVIDER. Missing credentials degrade cleanly instead of blocking API startup. Use make check-env for a secret-free summary and GET /api/v1/system/config-status for the same state from the API.

Replay setup requires only the checked-in fixtures and local PostgreSQL. Hosted Neon, Massive, World Monitor, and Alpaca credentials are not supplied by their source repositories. make check-live performs strict read-only checks; it does not create accounts, fetch repositories, or place orders. Alpaca paper execution requires ALPACA_PAPER=true (the MCP-compatible ALPACA_PAPER_TRADE=true alias is also accepted); LIVE execution is rejected.

The local graph integration is explicit: GRAPH_BACKEND=networkx is the safe default and GRAPH_FALLBACK_BACKEND=networkx is used if configured AURE paths are unavailable. AURE validation uses only the verified read-only command aure db check <path> --output json; it never initializes or mutates a graph database during make bootstrap.

Verification

make test
make lint
make typecheck
make build
make backtest-final
make paper-dry-run
make check-env
make bootstrap

The checked-in replay run is synthetic contract evidence: 18 trades, 1.2903% total return, 0.2182% max drawdown, 83.33% hit rate, and status INSUFFICIENT_DATA. It does not support a live-performance or final-holdout claim.

Demo and judge material

• Three-to-five-minute demo script
• Technical judge brief
• Judge Q&A
• Pitch outline
• Final test matrix
• Final audit
• Release manifest

Project structure

backend/       FastAPI, domain services, migrations, replay providers, tests
frontend/      Next.js UI, typed API client, accessible five-tab shell
configs/       Frozen strategy configuration
artifacts/     Replay, strategy, demo, audit, and release evidence
docs/          Contracts, architecture, runbooks, and audit records
scripts/       Safe validation and demo utilities