WHAT OUR OWN
RISK GUARDS
ACTUALLY COST.
A risk guard is a trade. It refuses an entry, and that refusal has a price: sometimes it saves you a loss, sometimes it costs you a winner. Every trading product sells you the first half. Almost none of them can tell you the second, because measuring it means admitting your own safety features are not free.
We measure it. Here is how, and what it currently says.
Every refusal is priced, not assumed
When a guard blocks an entry, the engine records the pair, the direction, the price, and — the part that matters — the stop distance the strategy would actually have used. That trade is then replayed against real prices as if it had been allowed.
Using the real stop is what makes the number honest. Replay a blocked trade with an arbitrary stop and you can make protection look like whatever you want. Replay it at the distance the sizer would genuinely have set, and you get the trade the engine would have placed.
Until this existed, every block was one line in a local log. Nobody could see what protection was worth — including us.
The result is a running scorecard. A negative figure means the blocked trades would have lost, so the guard saved you. A positive figure means they would have won, so the guard cost you.
| Guard | Entries refused | Replayed | Would have made | Verdict |
|---|---|---|---|---|
| Open risk guard | 10 | 2 | −27.6 pips | saved you |
| Spread guard | 2 | 2 | −21.9 pips | saved you |
| Symbol loss guard | 2 | 2 | −15.1 pips | saved you |
| Per-pair cap | 4 | 3 | +26.5 pips | cost you |
A live account, 30-day window. Replayed counts are lower than refusals because a trade only scores once its outcome is known. These samples are tiny — two or three trades per guard. They are shown as a mechanism working, not as a verdict on any guard.
What we did when a guard showed a cost
The per-pair cap was the only line reading positive. Twenty-six pips across three trades is far too small to conclude a guard is miscalibrated — and in fact one single GBPCHF trade accounted for +43.5 of it, so the average was one outlier wearing a trend’s clothing.
What it did do was prompt a look. And the look found something the number itself had not: the cap was counting resting orders as though they were open positions. On an account running a strategy that rests orders on twelve pairs by design and lets most of them expire, eleven of fourteen occupied slots held nothing at risk at all. Three pairs were completely full. Other strategies were locked out by orders that had never filled.
So the fix was not to loosen the cap. It was to correct what the cap counts: a resting order now costs half a slot rather than a whole one. Not zero — several limits genuinely can fill together, which is the stacking failure the guard was written after. At the shipped limit that still allows two open positions on a pair, or up to four resting orders, and it still refuses everything it refused before.
A measurement that looks bad is a reason to investigate, not a reason to act. The 26.5 pips did not justify changing a risk limit. The counting bug it led us to did. Had we simply relaxed the cap because the number was positive, we would have weakened a guard on the strength of one lucky trade.
The guard that fails at its own job
The weekend gap guard has a much stronger evidence base, because it can be tested against history rather than a handful of live refusals. A stop-loss is a trigger, not a guarantee — over the Friday close to Sunday open there is no trading at your level, so the fill is the first price on the other side of the hole. The guard exists to remove that exposure.
It has two halves, and over Feb 2023 – Jul 2026 they behave nothing alike:
| Half | What it does | Cost to the flagship | Effect on worst drawdown |
|---|---|---|---|
| No new entries | Nothing new opened from Friday evening to the Sunday reopen | under 2% of profit | none |
| Flatten open positions | Closes what the strategies already hold before the Friday close | about 10% of profit | slightly worse |
Read the second row again. A risk control that costs a tenth of the return would still be defensible if it bought a smaller drawdown. This one buys a bigger one, because closing at a fixed time exits losers that would have recovered.
We did not remove it, and we did not quietly leave it as it was. Three and a half years is roughly 180 weekends and contains no once-in-a-decade gap — so the flatten is real insurance against something this measurement cannot price. That is an argument for offering it. It is not an argument for recommending it, and it is certainly not an argument for implying it is free.
So the single switch became two switches, each carrying its measured cost in the app where you turn it on. The cheap half can now be taken without the expensive one, which was not previously possible.
While splitting them we discovered the flatten had never been reachable. The setting existed in the engine and was read by the guard, but nothing in the app could switch it on — while the guard’s own description promised it “can flatten before the weekend”. So no user ever paid that 10%, and the description had been over-promising since the day it was written. Both are fixed.
Guards you can see, tune, and overrule
There are twenty-six of them. Every one is visible in the app with a plain-English description of what it does and why, every one is yours to switch off, and every refusal it makes is replayed and scored on the panel where you would turn it off. Turning a guard off is a snooze rather than a silent hole: it comes back after a day unless you deliberately mark it permanently off, because the failure we were designing against was a user disabling everything while chasing a problem the guards were not causing.
That is the whole idea. Not “trust our risk management” — check it, on your own account, against your own refused trades.
The live scorecard rests on two or three replayed trades per guard, which is nowhere near enough to judge any single guard. The weekend figures come from a backtest over three and a half years, which is a real sample but still history rather than a forecast. We publish both at their true weight, and the direction of travel matters more than any individual number: when our own measurement says a safety feature is costing you, we say so.