Two Millennium pods made $3.7 billion. I reverse-engineered how.

Last month, two trading desks inside Millennium made about $3.7 billion.

Between them, that was more than half of the entire firm's profit for June.

And their primary strategy was index rebalancing.

One desk is run by Glen Scheinberg in New York. The other by Pratik Madhvani in Dubai.

Both trade the same boring thing: what happens when a stock gets added to or dropped from an index.

When a stock for example joins the S&P 500, every fund tracking that index has to buy it. On the same day, near the close, at whatever price the tape is showing. That is roughly $9 trillion of money that moves on a rulebook...

And the rulebook is public.

So the question is - if we know past patterns of how stocks are included, can we use AI to help us backtest what actually works for index rebalancing?

And that's where I had Claude Code reverse-engineer the whole thing. The exact rules the S&P committee follows, and 25 years of what the trade actually did.

Will this make you $3.7B? Probably not as there are more complexities to a full index rebalancing strat .... but index flows are too large for you to ignore no matter what strategies you trade.

Here is what we're doing:

• Rebuild the S&P 500 inclusion rulebook from the methodology
• Pull every index change since 2010 and line up the price action
• Measure what actually makes money, net of costs
• Turn the rulebook into a list of who gets added next

​

So here are the top 5 lessons on trading index rebalancing based on 25 years of backtests....

(Lesson 4 I ran AI to predict next adds...)

What I had Claude build

The engine reads the real eligibility gates first. A company needs an unadjusted market cap over $22.7 billion, enough public float, four straight quarters of positive GAAP profit, and a US home base. Miss one and you are out.

Those gates are why the game is winnable. A name does not get added because it is good. It gets added when a current member gets acquired or falls out, a seat opens, and the committee picks the best fit for it. More than 70% of new members are promoted up from the mid-cap index.

Then it pulled every S&P 500 change since 2010. That came to 659 events, of which 229 clean additions and 111 deletions had enough history to study.

For each one it ran a proper event study. Market-adjusted returns, lined up on the day the change takes effect, with the same stat tests a sell-side quant desk would run.

Here is the shape that came out. WSP - Index-Inclusion Arbitrage.pdf​

Lesson #1: The trade everyone knows is dead

The famous version is simple. A stock gets added, index funds have to buy it, so you buy ahead of them and sell into the forced bid.

It used to print money. In the 1990s an S&P addition ran about 7.4% on the news.

That edge is gone. Today it is closer to 1%, and the academic work that measured the decay could not tell it apart from zero.

Worse than gone, actually. Buy the addition now and hold it through the event, and you lose about negative 3.2%, net of cost. That is not noise. It clears the significance bar with room to spare.

Everyone crowded the obvious trade until there was nothing left. The buyers all show up early now, so the pop happens before the announcement you were waiting for.

If you are still running the textbook version, you are the exit liquidity.

Lesson #2: The edge moved to the reversal

Here is where the money actually went.

The passive funds still have to buy at the close on the effective date. They overpay for it. Then the stock mean-reverts.

Additions give back about negative 3.4% in the month after they join, and negative 6.3% over two months. The deletions do the mirror image. The names getting kicked out get oversold into the exit, then bounce positive 8.3% over the two months after.

These were the strongest and cleanest effects in the entire study. Everything before the event was statistical mush. The reversal was not.

The old trade was to chase the forced flow. The real trade is to fade it.

Lesson #3: The cleanest version needs no borrow

The obvious way to trade a reversal is to short the addition after it pops.

It works on paper, returning about positive 3.0% . But you have to borrow the stock to short it, and freshly added names are expensive to borrow. Push the borrow cost to 60% a year and the entire edge turns into a loss.

So flip the trade around.

Buy the deletion instead. The name getting dropped is already beaten down and oversold, and going long needs no borrow at all.

• Buy the deletion, hold a month: positive 2.8%, wins 61% of the time
• Short the addition: positive 3.0%, but only if you can borrow cheap
• Buy and hold the addition: negative 3.2% , the trap

The best trade in the whole study is the one a retail account can actually place. You are buying the unloved name everyone else is being forced to dump.

Lesson #4: You can back-solve who gets added next

This is the part that felt like cheating.

Because the rules are public, the candidates are computable.

So I pointed the engine at every mid-cap company and scored each one against the gates.

Here's the logic you want to bake into your prompt:

1. Scrape the index inclusion or drop rules (e.g., S&P500)
2. Scrape historical rebalancing events to discover any qualitative "tells" not on the hard criteria rules
3. Run a big list of stocks (e.g., Russell index or other list of names that are potential adds) to rank order most likely adds

I ran this for today and it surfaced the names that clear every bar based on the historical actions by the S&P500 committee: Twilio ($TWLO), MasTec, Carpenter, Illumina. Big enough, profitable, liquid, and US-based. The next addition is almost certainly sitting on this list.

Just as useful, it benches the imposters.

This is exactly why Marvell was so obvious. It was the largest eligible name in the hottest sector, waiting for a seat. Jensen Huang called it the next trillion-dollar company at Computex, a slot opened three days later, and it walked in. The stock ran about 57% into the date.

The committee still has the final say, so this is a probability, not a certainty. But you can systematically discover who may be up next.

Lesson #5: It is not free money

Here is the lesson that keeps you honest.

This same strategy lost Millennium about $900 million in early 2025.

The short legs die on borrow costs. The size is capped, because past a few days of volume you become the flow you were trying to trade. And the biggest, most certain moves are exactly the ones every other pod is already crowding.

A few percent of edge, run at real size with cheap financing across dozens of events a year. That is the business. It is a good one, and it is a volatile one.

So where does that leave you

Two moves.

Trade the mechanical events first. The Russell reconstitution and the Nasdaq rank changes are pure math with no committee, so they are the ones you can actually predict.

And here is where I would point the microscope next. When rebalance season comes around, take the candidate list and watch the options (somebody always knows something!).

Unusual call buying on one of these names weeks before the committee meets is a tell that a desk is already positioning. The flow tends to show up before the announcement does.

Personal

The more I use AI, the more I realize the work never actually stops.

Every time I finish something, I have three new ideas I want to run. The tools are so good now that the bottleneck isn't the work anymore, it's me. There's always another thesis, another build, another thing I could ship tonight. You end up doing more, not less.

A hundred years ago, when machines took over the factories, economists swore we'd all be working 15-hour weeks by now. Never happened. You just get more efficient and pour the time back into more output.

So productivity isn't the win. What you do with the time it hands back is.

This past month of the World Cup has been surreal (congrats to all my Spanish subscribers). The best game was England vs France, and I got to actually enjoy it while being 10x more productive than I was 6 months ago. All thanks to dozens of Claude Code and Codex agents running while I watched my favorite thing on earth: soccer (football, for everyone outside the US).

One of the biggest cheat codes with working with AI is separating the work that can run on its own from the work that actually needs you. Spend your working hours on the high-value tasks where you have to be in the loop, and let the agents handle the autonomous stuff while you're out living your life.

Anyway, back to it.