When Should You Turn Off a Trading Bot?

When it breaks a rule you wrote down before it started — not when it has a bad week. Every strategy has losing streaks, and the backtest tells you roughly how deep they should run. Turn the bot off when losses go meaningfully past that depth, or when the market changes in a way the strategy never assumed.

How do you tell a normal losing streak from a broken strategy?

By comparing what's happening to what the strategy's own history says should happen. A normal losing streak stays inside the range the backtest showed: similar depth, similar duration, similar market conditions. A broken strategy shows you something new — losses deeper than anything in testing, or losses arriving in conditions the strategy was never built for. A bot designed for a sideways, range-bound market isn't "on a losing streak" when a strong trend starts; its core assumption is simply no longer true. The question is never "is it losing?" — every strategy loses regularly. The question is "is it losing in a way its own history said to expect?"

What rules should you set before turning the bot on?

Three, at minimum, all written down while you're calm. First, a maximum drawdown line: a loss level — say, the backtest's worst drawdown plus a buffer — at which the bot gets switched off, no debate, no "one more day." Second, a review cadence: a fixed time, weekly or monthly, when you compare live results against the backtest — so you're judging on schedule instead of judging every red candle. Third, a list of the strategy's assumptions: which market conditions it needs, so you can retire it when those conditions visibly end. Pre-set rules beat panic for a simple reason: the rules were written by the calm version of you, and the switch-off decision always arrives when you're not.

What does the difference look like in numbers?

Suppose the backtest over three years showed a worst drawdown of 18%, with typical losing stretches of 6–10%. You set your shutdown line at 25% — tested worst plus a buffer for the future being unkinder than the past. Now the live bot is down 9%: uncomfortable, but squarely inside its historical behavior. Down 14%: worse than typical, still short of the tested worst — watch closely, change nothing. Down 25%: the line is crossed, the bot goes off, and you investigate before any restart. Notice what the numbers did — they turned "how do I feel about this?" into "which side of the line are we on?", which is the entire job of the framework.