Trusting the machine while it's in drawdown

The statistical version: you almost never have enough samples to distinguish a normal losing streak from a broken edge in real time. The decision has to be made in advance.

"My judgment is worst exactly when it feels most urgent" is going on a sticky note above my monitor.

shutdown number goes in the config, not your head. if its in your head you will renegotiate with it at 2am.

The part I'd add: the shutdown number is only half of it. Size it so that hitting that number doesn't wreck you, otherwise you'll turn it off early no matter what you wrote down. My pre-launch checklist has three lines, max drawdown to kill, max time flat before review, and what fraction of account this thing is allowed to touch. If I can't fill in all three I don't turn it on.

Genuine question for you. How did you pick the number? I've tried pulling the worst historical drawdown from the backtest and padding it, but that always feels like fitting to one sample of the past. Curious whether you use something like the distribution of drawdowns across simulated runs, or just a number you know you can stomach.

What helps me in those flat stretches is separating the two feelings, the loss itself, and the not-knowing. The second one is usually the heavier of the two. How do you sit with the waiting, Grizzly? Do you look at the curve daily or do you deliberately not look?

Dee — worst historical drawdown from one backtest is one draw from a distribution, so yes, padding it is guesswork. What I do instead: take the trade-by-trade returns and reshuffle the order a few thousand times. Same trades, same edge, different sequence. You get a distribution of max drawdowns and it's usually noticeably uglier than the one path you happened to backtest. Then set your kill number somewhere out in the tail of that, like the 95th percentile, and size so that number is survivable rather than sizing first and hoping. Two things worth knowing: expected max drawdown grows roughly with the square root of how long you run it, so a longer intended run needs a deeper tolerance, and it scales close to linearly with your bet fraction. Half the size, roughly half the drawdown. That last one is why Rita's point is the real lever.

Before you conclude the regime changed, check whether your fills drifted. I've had two "broken edge" scares that turned out to be wider spreads and more slippage on the same signals — the model was fine, the execution ate it. Log realised cost per trade against your backtest assumption and look at that alongside the equity curve. Cheap thing to rule out.