The Desk That Says No

I built a trading desk that spends most of its time telling me no. It is the most useful piece of software I own.

Some context. Day trading is one of my jobs. I trade small caps, mostly premarket and around the edges of the day job, and I do it from a machine that does nothing else. The desk is called Watchtower. It runs scanners, a chart, support and resistance levels, real time halts, my open positions, a trade journal, and a report that lands on my phone every night. It has exactly one hard rule, written into it on day one: it never places an order. It is a research desk. Every buy and every sell is still my thumb on my phone.

I did not write it alone, and I did not write it by hand. Two AI builders share the work. Claude Code and Codex both ship to the same machine, which is the single source of truth, and they leave each other notes in a shared project file so neither one undoes the other's work. Before either one changes anything, it has to compare against what is actually running. That rule exists because it got broken once.

The part I did not expect is that the desk tests its own ideas. One evening I had a panel of eight simulated experts review it: a broker, a hedge fund manager, an SEC staffer, a financial planner, a quant, a UI developer, a working day trader, and a seasoned investor. Everything they raised got built, tested, and deployed the same night. Some of what got tested failed, and the failures are the reason I trust it.

Here is what came back no.

Fibonacci levels. Traders swear by them. Watchtower tested them against real price action, about 3,300 tests per level. No edge. The toggle still exists on the chart, off by default, with a caption that literally reads "tested: no edge."

The pullback alert. I wanted this one badly. It was built, then graded only on sessions it had never seen. It did worse than picking a random entry on a stock that was already moving. It was rejected and it stays rejected.

The popular moving average. The one minute 9 EMA, a line half the internet draws on every chart, held as support worse than a random price. What actually held were plain printed prices: the swing low, the low and high of the day, whole dollars.

Then there was the quieter kind of no: the data itself was wrong in places, and nobody had noticed. Early morning relative volume was reading about twice too high, because the clock it used assumed volume arrives evenly through the day. It does not. Trades that ran into a trading halt were being scored at the price before the halt instead of where the stock reopened, which touched about one in ten alerts. A merge between two data sources was quietly blanking float numbers; 454 of them came back. For more than two weeks the nightly job stored every stock's daily range as a single flat price, so a level called "prior day high" was really reading the prior close.

None of those bugs threw an error. Every one of them made the screen look more confident than it should have been. That is the real risk in a tool like this. Not that it crashes, but that it is calmly wrong.

So the desk carries its own report card. Before I trust the nightly watch list, it has to earn it. After 30 graded sessions, the list must beat a plain baseline, the top ten gappers of the morning, with its whole confidence range above zero. At session eight it was behind: 12 percent against 21. The alerts, graded honestly including the ones that faded, win about 37 percent of the time across more than 4,000 of them. Those numbers print on the card every night, right next to the picks. I would rather look at an uncomfortable number every evening than a comfortable one that is not true.

What changed was me. The weekly review sorts every trade I take, and the first week's patterns were blunt. Trades before ten in the morning made money; trades after ten gave it back. Days when I stuck to a handful of names worked; days when I chased two dozen did not. The names the desk flagged first did better than everything I found on my own. After one bad Monday, the desk got a per trade loss cap: an alarm on my phone the moment a position is down more than the limit I set. After a stock reported earnings at eight in the morning and missed with nothing warning me, every name now carries a red earnings chip the night before.

None of this makes the desk a money machine, and I do not want this to read like one. It is a set of guardrails I built because I know exactly which mistakes I am likely to make, and I wanted something watching for them that does not get tired, excited, or hopeful.

The best feature is not a scanner. It is a system that will tell me no, with the evidence attached, and then keep telling me no until the evidence changes. I did not expect to build that. I am glad I did.

Nothing here is investment advice. Watchtower is a private research tool; it does not place trades, and no backtest promises anything about tomorrow.

MG