The most expensive click in my system

On a Monday morning I was logged into a server through a secure tunnel, looking at a virtual desktop, to untick a single checkbox. For the sixth time in just over three weeks.

That checkbox is called Read-Only API. When it is on, my system can watch the broker but cannot place orders. It is supposed to be off. In the configuration it is off. And still it flips back on after every fresh login.

That morning I knew for sure. This has to go.

What ran for four months

From 13 May to 18 September 2026 my system ran as a paper test at Interactive Brokers. Fake money, real prices. Every morning it scored 556 stocks on momentum, value, quality and catalyst, decided what should go in and out, and placed the orders itself. Nobody pressed a button.

The result: from € 50,009 to € 52,860. That is +5.70%. Over the same period the AEX did better and the S&P 500 did worse. The chart is on the dashboard.

To be fair: four months and one portfolio prove nothing. Not that it works, not that it doesn't. On top of that, trading days were missed because of outages, so this is not even the pure result of the strategy.

Three numbers, one of them honest

When I saw the final result, I wanted to put a different number on the site. The positions still in the portfolio at the close were up +9.5% together. That sounds a lot better than 5.7.

It just isn't a return. My system sells losers through a stop loss or a trailing stop and holds on to winners. So whatever is left at the end always looks better than the whole. The 33 trades that were closed along the way lost money on balance. Look only at the basket that remains and you leave out half the story.

There was a fair point behind my reflex, though. On average 21% of the money sat in cash. That is not an accident, it is design: there is a cash buffer, and when markets get nervous the system buys less. Take the full result, including the losing trades, and set it against the part that was actually invested, and you get +6.97%.

All three are on the dashboard now, each with a note on what it means. The headline stays 5.70%. That is what happened to the whole amount.

Where the time really went

Not into the strategy. Not into the factors, the backtests or the website. Most of the problems, and most of my time and attention, sat in one component: the connection to the broker.

Placing an order sounds like one step. In practice it is writing the order, reading back whether it arrived, checking that the status is right, waiting for the execution, matching that execution to the right order, updating the position, and every morning verifying that the broker's books and mine still say the same thing. Each of those steps has its own exceptions. An order that was placed but never confirmed. A number that gets handed out again after a restart. A connection that is alive but no longer allowed to do anything.

Underneath all that runs a program that was really made for a desktop. It has to stay logged in on a server, it restarts itself every night, and for real money it needs two-factor authentication as well. It feels like technology from another era that you have to fold a modern system around.

And now the number that puts it all in perspective: my system makes on average one trade every four trading days.

For that one trade I had a server, a tunnel, a watchdog script that restarts things, a set of alarms and a playbook for when the checkbox flipped again. The complexity was out of all proportion to what it delivered.

What replaces it

The system keeps deciding fully automatically. That part works, and that is where the value is. What changes is the last mile.

Instead of an order to the broker, a message goes to Telegram. What to buy or sell, how many shares, at which reference price, and why: which trigger, which scores. I then place the order with my broker myself. The next morning the system books the trade at the opening price, and if reality went differently, I correct it.

There will be two tracked portfolios, one of 50k and one of 10k. They do not move in lockstep: the small one holds eight positions instead of fourteen, so it gets different signals. Both are tracked in public.

One more thing that needs saying. In earlier posts I wrote that the system would go live with my own money from June. That did not happen. So far only fake money has run. I am leaving those posts as they are, they describe what the plan was at the time. Real money follows only once the new setup has proven to run reliably.

I am giving something up too. Executing by hand takes discipline. I can skip a signal, be late, or hesitate. The system is going to record that, so it becomes visible what my own behaviour costs. That will probably be the most interesting measurement of all.

What I take from this

With AI you can have almost anything built. That shifts the question. No longer: can it be done? But: is it worth maintaining?

Automate the thinking. That is where a system beats me: the same rules every day, no emotion, no exceptions. Automating the clicking only makes sense if you click often enough. Once every four days is not often enough to keep an old API alive for.

The most expensive click in my system was not an order. It was that checkbox.