One stock is the whole lead

On April 2 I handed an AI an investing account and set it against the S&P 500.

Six weeks later I paid $1,050.03 for a year of market data, because I'd worked out I wasn't testing the AI. I was testing whatever it could scrape that morning.

Five months on, I still can't tell you whether it works.

That's not me stalling. That's the result, and it cost a thousand dollars to get.

I'm not a financial advisor. Nothing here is a recommendation. I'm naming every position below and putting the tracker online so you can check my arithmetic, and I want to be blunt about why: it's so you can judge the result, not so you can copy it. If you're thinking about doing this yourself, talk to a fee-only planner before you talk to a chatbot.

Everyone has an opinion about whether AI can pick stocks. Opinions are free. That's why there are so many of them. I wanted the version that costs something.

Real money. $5,000 to start, plus two $100 deposits a month. Not a paper account, not a backtest, not a screenshot of a simulation.

The AI makes every decision. It never makes a trade.

That distinction is the whole design, and I had it muddled in my own head for the first month. The machine writes instructions. I read them and place the orders myself. No API key, no account access, no path from the model to my money. Not because I don't trust it. Because a system that can move its own cash isn't an experiment, it's a liability.

Three things it can't change, ever, without me signing off: the deposit schedule, the cap on asking for extra money, and the rule that it never executes. Its own strategy document says why, and it's the best sentence in the whole setup: the AI must not be able to expand its own claim on my cash.

The benchmark is the S&P 500. Here's the part most people posting their returns skip. My money is Canadian. Unless you convert both sides to the same currency before you compare them, part of the gap you're looking at is the loonie moving, not the machine working.

If I'm going to tell you the result means nothing, I should show you why it's still worth running.

It sees what you see. Public filings, public prices, public news, plus PDFs I remember to upload by hand. No live feed, no order book, no research terminal, no company access. It can't even read the account. It knows what a screenshot tells it, and on days I don't send one it says so in the file rather than guessing.

It runs on a clock, not on impulse. Seven scheduled jobs: a daily monitor, a weekly deep analysis, a weekly earnings sweep, a contribution allocator on the 1st and 15th, a monthly scan, a monthly sell review, a monthly audit. That's about 529 decision points a year. In the first sixteen weeks it produced 19 trades.

That ratio is the point. Most days the right answer is do nothing, and the daily job is built to stay silent when nothing fires. Of 249 logged entries, the most common outcome is a variation on nothing happened.

Everything traces to a numbered rule. Five scoring dimensions, zero to two each. Three hard floors that block a purchase no matter how good the total looks. Eight sell triggers, each with a pre-committed response time — one trading day for a broken thesis, same day for an emergency. A 30-day minimum hold. A 30-day cool-down before it can rebuy anything it just sold.

And it can ask me for extra money, capped at $500 per rolling thirty days and $2,500 a year, measured on cash actually deployed. It's used $292.96 of that in five months.

One thing I'll flag because it cuts against me. Since August the AI amends its own strategy document, on the theory that me editing the rules would contaminate the test. It's gone from v2.0 to v2.8 in under four months. Nine amendments. Which means the strategy under test hasn't held still for a single full quarter, and that damages clean attribution just as badly as the short sample does.

Nine positions. Everything, with weights, as of September 16.

Nine closed trades too, for +$240.47 realized. Four of them lost money.

The whole tracker is online and you can look at it yourself: portfolio-tracker-ten-eta.vercel.app, password guest. It's read-only, it's the same dashboard I look at, and it updates when I update it. I'd rather hand you the numbers than ask you to trust my summary of them.

My own dashboard gives me more than one answer, and that's the most useful thing on it.

Gain on the money I've put in: +18.0%, which is +$358 ahead of the index, or +4.9%.

Time-weighted return, which is the fairer comparison because it doesn't reward me for adding cash: +23.1% against the index's +15.7%. A lead of 7.4 points, not 4.9.

Both are honest. Both come off the same screen, over the same five months. They disagree by two and a half points of margin, and that's before anyone asks whether five months means anything.

The system's own headline metric is neither of those. It runs a shadow ledger. Every real deposit gets simulated buying the Canadian-listed S&P 500 ETF on the same day. Same cash, same dates, no way to flatter it by choosing a start date. When it can't measure both sides on matching dates, it says so in the file instead of quietly picking the flattering pair.

Here's the number that should stop you from being impressed.

HALO has made +$431 of the gain. The entire lead over the index is +$358.

One position out of nine is worth more than the whole margin. Take it out and this AI is behind the S&P 500. Add the second-best name and two holdings account for about two-thirds of everything the account has made.

That's not a portfolio beating an index. That's one stock going up 51.5% while eight others come along for the ride.

And it wasn't free. The worst the account has been underwater is −9.9%, back in July. The index's worst stretch in the same window was −3.8%. So the lead came with roughly two and a half times the drawdown, which is a thing almost nobody mentions when they post a green number.

This is the part I didn't expect to be the most interesting, and it's the reason I still think the thousand dollars was well spent.

In August the system put $100 into a company whose quality score was openly in dispute at the time. The dispute didn't get resolved first. Sixteen days later the score failed properly and the rules forced a full exit at a −9.09% loss.

Read that again. The biggest realized loss in five months didn't come from a bad stock pick. It came from a process gap. Money went into a name whose eligibility was still an open question.

There's more. A misplaced write overwrote eight of ten position rows in August, and it took two days to notice. About eight scheduled runs never fired in July, including both month-end days, because the jobs only run when the app is open. A stale summary row hid a position-size breach for two weeks. And for a stretch one position was either 14.24% or 15.90% of the book against a 15% cap, depending on which denominator you used. Nobody had ever defined which one was right. Two different jobs were quietly using different ones.

Every one of those is in the log because the system is built to log it. If you don't log it, it didn't happen. Each one produced a rule, and the rules are numbered, dated, and attached to the incident that caused them.

That's the actual finding at five months. Not that a machine can pick stocks. That a controlled process makes its own failures visible instead of invisible, and visible failures are the only kind you can fix.

In July I wrote about the flat part of a compound curve, and how most people quit standing right there, months in, when the graph still looks broken. That's the famous mistake and I've made it plenty.

This is the other one. When the line does point up, believing it.

If this thing is still green at Christmas I will absolutely want to write you a piece about how the machine has it figured out. Same short window. Same absence of evidence. Opposite conclusion, and a much more fun one to publish.

Quitting early and calling it early are the same error. One of them just feels better.

So: five months, nine positions, one stock carrying the margin, a strategy document that's been rewritten nine times, and three defensible ways to measure it that don't agree. Anyone who tells you that adds up to a verdict is selling something.

Almost all of my money sits somewhere else, and it does none of this.

Automatic deposits into index funds, on a schedule I set once and have barely touched since. Nothing timed, nothing picked. I published the whole boring table back in June. Across the same five months I haven't logged in to make a single call on any of it.

That isn't laziness. It's a conclusion I came to years ago: I wasn't going to beat the market by paying closer attention, so I stopped paying attention.

Which should make this experiment the last thing I'd ever run.

Here's how I square it. Both accounts start from the same place, which is that I don't trust my own stock picking. The index account acts on that by deleting the decision. This one acts on it by asking whether something that isn't me can beat deleting the decision. Same distrust, pointed two different ways.

One of them is where my retirement actually lives, and it'll keep running long after I'm bored of this. The other is a $1,050.03 question I wanted a real answer to instead of a take.

For the record: I check it out of curiosity, nothing more. My assumption from day one is that I'll lose all of it, so I'm not too worried about the money. And I've never once wanted to override it. An override would invalidate the results, and the results are the only thing I'm paying for.

Ask me again in two years. That's when the number starts meaning anything, and I'll publish it either way, including the version where the account I never touch beats the one I paid a thousand dollars to build.

Next week: the race entry I bought in May, and what I did about it.

Forward this to one person who's already sure they know the answer.

— Kiegan