Paper Trading

The same deployment, no capital — a mode, not a separate product.

Paper trading is "mode": "paper" on POST /runtime/deployments. Not a different resource, not different endpoints — the same deployment object, running the same code on the same runtime against the venue's live prices, with simulated fills and a simulated balance.

Launch scope: paper mode ships with the freqtrade venues (Hyperliquid, Binance), where it runs on freqtrade's own battle-tested dry-run engine. Nautilus venues (Polymarket, Lighter) launch live-only; wiring Nautilus's sim execution into the tenant runtime is planned and tracked in the design log. /context/venues reports paper_supported per venue — check it, don't assume.

json
{
  "framework": "freqtrade",
  "venue": "hyperliquid",
  "mode": "paper",
  "name": "btc-range-rehearsal",
  "code": "…",
  "config": { "timeframe": "4h", "dry_run_wallet": 1000 }
}

What changes in paper mode:

  • No credentials. credentials is ignored; there is nothing to sign with and nothing at risk.
  • No funding requirement. Start succeeds on an empty venue account.
  • Simulated fills. The framework's own dry-run engine (freqtrade dry_run today; Nautilus sim execution when it lands) fills against live market data.

Everything else is identical: logs, metrics, lifecycle, alive_until, isolation. GET /runtime/deployments?mode=paper filters the fleet.

The intended loop

  1. Backtest — does the idea survive history?
  2. Paper — does it survive now: live data, real latency, real signal cadence, zero risk?
  3. Live — same payload, "mode": "live", credentials attached.

Paper is the step that catches what backtests structurally can't: a strategy tuned to a regime that just ended, an entry condition that never fires on current volatility, config that references a delisted pair. It costs nothing and runs unattended — there's no reason to skip it.

Simulation limits

Simulated fills are optimistic: no real slippage competition, no partial fills, no funding payments on most sim engines. Paper PnL is an upper bound on live PnL, not an estimate of it. Metrics responses for paper deployments carry "simulated": true so downstream tooling can never mistake rehearsal numbers for real ones.

Previous
Backtests