Where the cliff comes from
Moving from research to a running bot usually means starting over.
- The strategy is rewritten in a bot framework, and subtle differences creep in: a candle boundary, a rounding rule, an order type.
- Webhooks, API keys and order plumbing have to be wired, each a new place for a bug.
- Position sizing and leverage maths are redone by hand.
- Nothing sits between “it backtested well” and real money.
A sane path to live
- 1. Backtest after costs, and read the result by market condition, not only the average.
- 2. Validate on data the search never saw: a holdout or walk-forward windows.
- 3. Paper-trade the same rules on live prices with no real orders.
- 4. Compare paper with the backtest: entries, exits, trade count, and the gap between expected and actual fills.
- 5. Start live small, with limits you can afford to lose entirely, and watch it.
- 6. Keep a stop-the-bot rule you decided on in advance, not in the middle of a drawdown.
What to compare on paper
A paper run will not match the backtest exactly, and it should not. Look for differences you can explain: a trade the backtest took that paper skipped, or fills at slightly worse prices. Differences you cannot explain are the ones to chase before any money is involved.
What to watch once it runs
Decide the warning signs before the strategy is running, so you are not inventing them mid-drawdown.
- Trade frequency. A strategy that suddenly trades far more or far less than its backtest did has changed, or the market has.
- Slippage and fees against the assumptions. If real costs are larger than the model's, the edge shrinks with them.
- Funding paid or received, for positions held across funding times.
- Drawdown against the backtest's worst. A live drawdown beyond anything the backtest saw is a signal to stop and look.
- A written stop rule. The point at which you pause the strategy, decided in advance.
How the Lab closes the gap
A deployment pins a version of your strategy and runs it on the Lab's own runtime, using the same decision and position code as the backtest, on a live Binance kline feed. Nothing is rewritten, so the strategy you tested is the strategy that runs. Paper starts in one click and never touches an account.
Live goes through an exchange adapter you set up, and needs your explicit approval in a signed-in browser. A token, and therefore an AI agent, can manage paper deployments but cannot approve a live one. See paper trading.