An alert count is not a workload and it is not a trade count. A rule that fires eleven times a week can produce one position or nine, and the difference is not cosmetic. It decides how much cash you need sitting uncommitted, how many round trips of commission you pay, how often you have to be at a screen, and whether the strategy you believe you are running is the strategy your filters actually permit.
The number that closes that gap is a conversion rate. Of the signals that arrived in a window, what share cleared every filter, sizing rule and risk guard and became a position. Autopilot's Backtest tab gives you the numerator for one profile over one window. The denominator is yours to build, and the honest version of this exercise is mostly bookkeeping.
Cadence is the number you can plan around
Think about what you would do differently if you knew a profile produced two trades a week rather than eight. At two, you can look at every entry the evening it happens, you can hold most of your account in reserve, and a fifteen dollar round trip is noise. At eight, you cannot review them individually without giving up your evenings, your reserve gets drawn down in clusters, and that same fifteen dollars is now a hundred and twenty a week against whatever the strategy earns.
Alert volume tells you none of this, because the filters sit between the alert and the order. A profile that requires a signal plus available exposure plus a position count below its cap will decline a large share of what reaches it, and the share is not stable. It depends on how bunched the signals are, which depends on the market. Two profiles fed by the same rule can convert at wildly different rates purely because one of them is usually already full.
The three inputs the Backtest tab actually takes
The Backtest a Profile panel is deliberately small. It replays historical signals against a chosen profile with no real orders placed, and it states that it honors all filters, sizing, leverage caps and risk guards. That last clause is the whole reason this measurement works: the replay is not asking whether the signal was good, it is asking whether your configuration would have let you act on it.
You give it three things. A profile, from the list of profiles you have configured. A number of days. And a start equity. Set the day count to ninety and you have your quarter. Set the start equity to the number you would genuinely have committed to that profile, not a round figure that flatters it, because equity feeds sizing and sizing feeds every cap in the chain.
Whatever the run reports back, the only figure you need from it is a count of trades taken over that window, alongside the window itself. Write both down. If the run does not hand you a total directly, count the rows. This is a two column spreadsheet, not a research project, and its value comes from repeating it rather than from precision on any single run.
Building the denominator you do not get for free
The count of signals that fired is the harder half, and pretending otherwise is how this exercise goes wrong. The Alerts ledger shows each rule with a last fired date, which tells you a rule is alive but not how often it spoke. The History tab alongside the main alerts list is where the record of firings lives, and it is worth opening before you start inventing a denominator.
Three approaches, in descending order of how much I trust them. Best is a real count of firings over the same ninety days from the history record, filtered to the source that feeds this profile. Second is your own tally, which is only available if you started keeping one, and which is the reason to start today rather than when you next want this number. Worst, and still better than nothing, is a signal count taken from the source engine itself over that window.
Whichever you use, the denominator has to be the same universe as the numerator. If the profile follows one instrument on one timeframe, count only the signals for that instrument on that timeframe. A conversion rate computed against every alert you own is not a conversion rate, it is a ratio of two unrelated numbers, and it will be small enough to look alarming for no reason.
Reading a rate of one in eight
Suppose ninety days produced forty signals and eleven trades. That is roughly one in four, about one trade a week, and it is a comfortable number: you can review each one, and a cluster of three in a day is unusual enough to notice.
Now suppose the same forty signals produced five trades. One in eight. The interesting question is which constraint ate the other thirty five, and there are only two families of answer. Either an entry filter rejected the signal on its own merits, which is the filter doing what you designed it to do, or a portfolio level cap was binding, which means the profile was already full and the specific signal was irrelevant. These call for opposite responses. A selective filter is working. A cap that is permanently binding is not a risk control, it is a queue, and it means the first signal of every cluster gets taken and the rest are discarded on the basis of arrival order rather than quality.
You can separate the two by re running the replay on a duplicate profile with a single cap widened far enough that it cannot bind, holding everything else identical and running both on the same day so the day count covers the same calendar. If the trade count barely moves, your filters are doing the selecting. If it jumps, your caps were the real strategy.
One caution on the arithmetic: a rate measured over ninety days of one regime is a description of that regime. A quarter where the market trended and signals arrived evenly will convert very differently from a quarter where everything fired in the same three days. Run it for a few different ninety day windows before you treat any single figure as your cadence, and expect the spread between them to be wider than you want.
The gap between a replay and a Tuesday morning
Two things separate this measurement from what will actually happen, and both of them point the same way, which is toward optimism.
The replay prices entries at the signal, and the module says so plainly on this screen: 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 you are solely responsible for the trades. It tells you to start in paper mode, verify the behaviour, then go live with small sizes. A conversion rate is one of the specific behaviours worth verifying that way, because a live profile can also fail to convert for reasons the replay never sees, such as a rejected order or insufficient balance at the moment the signal landed.
The second gap is the one that matters for safety. Knowing your cadence tells you how often the machine acts without you, and that is the point at which the guardrails stop being optional. A profile you have measured at one trade a week is not therefore a profile you can leave alone for a week, because the measurement is an average and the failures cluster. The controls that stop things are the profile's own risk guards, the pause action, and the emergency kill switch on the Autopilot page, and that kill switch is a button somebody presses. It is fleet wide and it is manual. Nothing on this page is watching your conversion rate and halting when it drifts, so if your plan for a profile that suddenly starts taking six trades a day is that you will notice, write down how you will notice, and put a hard cap on the profile in the meantime.