What the banker does on its own — and where it has been wrong

You state a mandate once and it builds the portfolio, watches it, prices the rebalance and prints what your choice costs. Here is exactly what it does between your visits, what it will not do at any price, and the part most product notes leave out: the calls it got wrong.

Dear reader,

The hardest part of a rules-based portfolio is not the rules. It is the Thursday you do not feel like running them. So we built the part that runs them for you, and we gave it a job description rather than a personality: it is a private banker, its subject is portfolio construction, and it is meant to be handed the whole of it — take a profile, build the allocation, rebalance it, and explain what the choice costs.

You can drive all of this yourself in the demo preview on simulated data, without an account. Figures quoted below come from that demo — simulated data · illustration only. This note is the written version.

One question, and what each answer costs

A first-time visitor is asked exactly one thing — what are you here to do? — with six concrete answers. Picking one sets the weights, the horizon, the drift tolerance, the guard it argues hardest and the standing mandate it then works to. You configure nothing.

The part we care about is what comes next. Every answer is priced, from the model’s own record, before you commit to it. Answer “Protect what I already have” and it says so plainly:

Safety is bought, not free. These weights give up compounding in exchange for a shallower fall. In a rally you will lag, and you will notice.

Answer “Compound this for decades” and the bill arrives from the other direction:

These weights fall noticeably deeper at their worst than the conservative mix — a loss you would have to sit through without selling.

And the middle answer, “Stay ahead of prices”, is told it will never win either argument: some compounding given up against the growth weights, and a deeper fall than the conservative ones. Presenting a risk profile as a free choice is the oldest trick in this industry. We would rather print the invoice next to the promise.

What it does while you are not there

Once a mandate is standing, the banker works on its own — on the clock and on every change you make to the profile, horizon, regime or your own sleeves — and writes every action down with what it did, why, and the number it produced. A sequence from a live session on the balanced mandate, on €250,000 at this week’s risk-off reading of 38:

Two things about that log are deliberate. The first is that most entries say a version of nothing changed — that is the honest output of a slow model, and it is written down as plainly as the weeks it acted. The second is that the last row is not a marketing exercise: auditing its own record is one of the nine sweeps it runs, on the same rotation as everything else.

What it will not do, at any price

Ask it for leverage, margin, hour-by-hour trading or a return you can count on, and it does not soften the answer:

No — and not a milder version of it either. This model is unleveraged, it rebalances once a week rather than by the hour, and every tilt is sized at a quarter of the mathematically optimal bet. I will not propose borrowed money, hour-by-hour trading, or a return you can count on.

Ask it to act on something a reader argued well for in the forum, and it refuses that too:

I will not move a weighting because a forum post argued for it: the sleeves come from the regime model and the published bands, and nothing else gets a vote. If a thread contains a claim you want tested, name the series in it and I will read it against the record.

And it will not place the trade. It prices the rebalance in full — trades, euro moved, expected slippage, and which legs it skipped on the 20 bps cap — and then stops, because that is where a published model ends and a decision about your own money begins. If you change the profile out from under your own mandate it raises the mismatch and still leaves the call with you: “I will not move it back without being told to — you changed it, so you decide.”

Where it has been wrong

This is the section every product note skips. Simulated data · illustration only. Over the 152-month illustrative sample in the demo, on the balanced weights:

• 4 of 14 completed directional calls wrong. The costliest was the risk-on call at 2014-03. The wrong ones are on the free tier, printed at the same size as the right ones.
• Behind a plain 60/40 since 2023-10, 15.3 points off its own high-water mark, and it has not been made back. In the product’s own words: “A client who joined at that mark has been paying for insurance they have not yet needed.”
• Ahead in barely half of individual months. Close to a coin flip.
• The sample is too small to rank. Ask it whether the hit rate proves anything and it says: “No, and I will not dress it up. 14 completed directional calls is a small number. A coin flip on 14 tries has a standard error of about 13 points, so a hit rate on a sample this small is not something you can tell apart from luck.”

What the record is evidence for, at most, is the shape of the outcomes rather than the accuracy of any single call: shallower losses than a static 60/40 on the same simulated data, with the −20 % drawdown halt never firing over the sample. That is a claim about losses, and it is the only one we are willing to make.

How to stop it

Pause is one click, and it stops checking and stops changing anything — “Nothing already in the log is removed — a log you can quietly edit is not a log.” Undo reverses its last change and leaves the original entry struck through. Changing the mandate reopens all six with their costs, and keeps the log. Clearing it entirely leaves the banker idle, and it says so rather than pretending to be busy.