A stop is a level at which you have asked to transact. It is not a level at which you will transact, and the distance between those two statements is the entire subject of unattended automation. For a profile that holds overnight, the loss you actually take on a bad morning is set by where the market reopens, and no configuration on the platform has any influence over that number.
Most desks know this and still size as though the stop were binding, because the backtest reports a clean exit at the level and nothing in the output flags which of those exits were fantasies. The correction is to partition the simulated record by whether a human could have acted, and then to size the profile off the part where they could not.
Attended is a property of your rota, not of the exchange
The first thing to fix is the definition. Unattended does not mean the market is closed. It means nobody on your desk was watching, reachable, and authorised to act, and those three conditions fail at different hours.
Look at what the profile list on this desk actually spans. Several equity names on a daily cadence, an index profile, a silver pair and a crypto pair on four hour cadences. A daily cadence equity profile holds through every overnight session by construction: the position exists precisely because the signal arrived at one daily bar and will be reconsidered at the next. It is unattended for roughly seventeen hours out of every twenty four before you even reach the weekend. The four hour crypto profile never closes, which sounds like the opposite problem and is in fact the same one dressed differently, since your coverage still stops at six in the evening.
Write the coverage calendar down as hours, per profile, and put an owner against each block. Then compute the fraction of the profile's holding time that falls outside it. On a daily cadence profile with a business hours desk, that fraction is usually above seventy percent, and the number is worth having in writing because it reframes every subsequent conversation. You are not running a supervised strategy with some overnight exposure. You are running an unsupervised strategy that you occasionally watch.
Splitting the simulated record into three states
The Backtest a Profile panel replays historical signals against the profile with no real orders, honoring the profile's filters, sizing, leverage caps and risk guards, and takes a profile, a day count and a start equity. That gives you a trade record under your own constraints. I cannot confirm that it breaks results down by session or by time of day, so plan on doing the partition yourself from the trade list timestamps.
Sort every simulated trade into one of three states based on when the loss accrued rather than when the trade opened. Attended, meaning inside your coverage window with the instrument trading. Unattended and tradeable, meaning outside coverage but the market was open and someone could in principle have been woken up. And unattended and untradeable, meaning the market was closed, halted, or limit locked, so intervention was impossible regardless of staffing.
Then report three numbers per state: share of total profit and loss, share of the worst decile of outcomes, and the largest single loss. The pattern that shows up most often is that the attended state holds a modest share of the profit and almost none of the tail, which is uncomfortable reading and exactly the argument you need when someone proposes that overnight risk is managed by the stop.
The third state deserves its own line in every risk report, because it is the only one where no amount of process fixes anything. Staffing addresses the second state. Sizing is the only lever on the third.
Sizing on the gap rather than on the stop
Once you know which trades lived in the untradeable state, size for them explicitly. The method is to replace the stop distance in your position sizing formula with a gap estimate for that instrument, and to do it for any position that will be held across a closure.
Build the gap estimate empirically rather than assuming a multiple. For each instrument in the profile, take the distribution of overnight moves from prior close to next open across as long a history as you have, and separate the weekend closures from the weeknight ones, since the weekend distribution has a fatter tail for no reason more complicated than more elapsed time and more scheduled news. Pull the 99th percentile adverse move. That number, not your stop, is the loss you should assume when sizing a position that will be open through a closure.
The consequence is usually that the overnight position is materially smaller than the intraday one, and that is the honest answer rather than a failure of the strategy. If the profile cannot carry meaningful size once sized this way, you have learned something useful about its capacity, and the alternatives are a flat-by-close rule, a hedge that survives the closure, or accepting the exposure with a documented number attached. All three are defensible. Sizing on the stop and calling it risk managed is not.
Two adjustments are worth making. Scheduled events, earnings above all, should be handled by a calendar filter rather than by a gap percentile, because the distribution conditional on an earnings release is a different distribution and averaging them understates both. And correlated gaps matter more than individual ones: several equity profiles built on the same signal source will frequently be positioned the same way into the same macro release, so the desk level overnight exposure is the number to size against, not the per profile figure.
The instrument that never gaps carries more unattended risk, not less
A 24 hour crypto pair looks like it solves the problem. The price series is continuous, so there is no reopening gap in the conventional sense, and a stop can in principle execute at any hour.
The exposure simply relocates. Continuous trading means the position is live through every hour your desk is dark, and liquidity in the small hours of a weekend is a fraction of what it is on a Tuesday afternoon, so a move that would be absorbed in normal conditions travels further. The stop executes, at a price that reflects an order book nobody is supporting. In practice this produces a slippage distribution on unattended fills that is materially worse than the attended one, and it will not appear anywhere in a replay, because the replay places no orders at all.
Measure it once live. Tag every fill with the hour it occurred, compare the realised price against the signal price, and split the distribution by coverage state. If unattended fills are consistently worse, that difference is a real cost of running the profile around the clock and belongs in the profile's expected return, not in a footnote about market conditions.
What the halt does by the time you reach it
The platform is direct about the operating risk on this page. With paper mode off, Autopilot places real orders on connected exchanges, execution prices may differ from signal prices because of market conditions and latency, and the user is solely responsible for every trade. The recommended sequence is paper first, verify the behaviour, then live at small size. The controls offered at the fleet level are pause all, resume all and an emergency kill switch, and all three are buttons that a person presses.
Which means the response time on an overnight event is your escalation path, not your software. Write down who is reachable at three in the morning, through what channel, with what authority, and how long it takes them to reach a terminal. Then be honest about the fact that by the time they arrive, a gap has already happened. The kill switch prevents the second bad trade and the third, and it prevents a malfunctioning profile from compounding an error across a whole session. It does nothing whatsoever about the first loss, and a risk framework that treats it as a stop for overnight exposure is describing a control it does not have.