In this issue:
-
Taking enough risk
-
ANT, fwiw
-
Making sense of returns
-
Statistically significant
Friends,
Before we get to the Money Angle meat, I want to share a thought that’s been swinging through the vines in my skull.
A little background.
About every 6 weeks for the past 5 years or so I grab dinner with 2 close friends at the Hillstone in SF. The restaurant fits like an old t-shirt from my NYC days when it went by Houston’s. This regular dinner is one of my most cherished rituals. The 2 friends have traded options for longer than I have at this point. One is still very much in the trenches, he has too many kids to ever retire. The other retired from a long career that focused on the SPX and VIX options complex. Today he runs the college admissions service that my family and friends either use or will be using.
You know how dinners like this go if you’re my age. Some time reminiscing, some time solving the world’s problems, which includes everything from markets, politics, what’s going on under the hood in single stock vs index vol surfaces, current AI workflows, the future of our children, and always the shitshow of anything related to college. (There’s also a healthy dose of sports and what concerts we’re eyeing).
It’s a GenX salon with the deviled eggs followed by a kale chicken Caesar and washed down with an NA pilsner (for me anyway…the bill confesses wine for them). Setting aside my dream of starting a holding company for all our dumb ideas, Hillstone LTD, I always come away from these dinners ruminating.
The recent one was no different.
When I got to the restaurant, while waiting to get seated it was just me and the current trader who was also my biz partner for most of my pro days. You spend about 15 years arguing over trades 20x a day, and whether you take my word for it or you violently agree because the nature of your work leaves you mentally bare to a trusted colleague, you know each other better than you know your spouse. I forget the trigger, but I said something in a passing way while making a wider point. The passing comment was “we should have taken way more risk”.
Even though we never really explicitly discussed that, he replied in a way that indicated that he thought that was true with the same certainty that one has about the sky being blue. He didn’t pause and say, “what do you mean, we didn’t take enough risk?”.
With the benefit of time, this was just so obvious to us that the fact blends in with the rug. Not able to let the obvious be obvious I had to question if this was just survivorship bias but I think where we landed on this is just another version of something I wrote in A Former Market Maker’s Perception of PFOF describing what I think set SIG apart from many of the firms on the floor:
It started with skill and luck. The early big bets on talent and technology meant they were bringing guns to a knife fight. SIG wasn’t known as the “evil empire” on the Amex just because of the black jackets we wore. They understood the risk-reward was completely outsized to what it should be 25 years ago. They were amongst the first to tighten markets to steal market share. They accepted slightly worse risk-reward per trade but for way more absolute dollars. They then used the cash to scale more broadly. This allowed them to “get a look on everything”. Which means you can price and hedge even tighter. Which means you can reinvest at an even faster rate. Now you are blowing away less coordinated competitors who were quite content to earn their hundreds of percent a year and retire early once the markets got too tight for them to compete.
SIG was playing the long game. The parallels to big tech write themselves. A few firms who bet big on the right markets start printing cash. This kicks off the flywheel:
Provide better product –> increase market share –> harvest proprietary data. Circle back to start.
The lead over your competitors compounds. Competitors die off. They call you a monopoly.
In short, SIG understood that bookies could get rich on much less edge than they were getting on trades at a time when volumes were exploding (late 1990s), but the broader investing world hadn’t caught on to just how lucrative the risk-reward was. Nothing attracts a crowd like ROE, of course, so it would be a fleeting period, but SIG maximized.
I think my friend and I, seeing how much tighter markets have gotten, to the point where Jane Street, with its massive balance sheet and coverage, could sensibly price risk with a linear utility of wealth, and just think of how we’d kick a grandma down a flight of stairs for a penny of edge when we used to make a half-dollar.
But this progression of efficiency is good. But if you find yourself riding an outsize wave, you should make sure to check in with yourself…Am I taking enough risk?
I don’t mean this in the sense of investments in public assets. I’m talking about edges you find that others haven’t caught up to yet. New ways of doing things that you wouldn’t want to shout about. Stuff that is in the gears of your business or industry as opposed to wagers open to everyone. And once you find a loop with an attractive return that can take capital, you push it because eventually your little secret isn’t so secret. Eventually you’re drawing blood from a stone, crushing your learning rate and, if you’re slow to adapt, your sense of hope and growth.
I hope these thoughts feel timely. I don’t know if AI is gonna eradicate us, but it’s simultaneously killing some models and opening up new ones. Get to the right side of the K.
Speaking of AI, ANT went public Thursday and closed about 17% above its IPO price. A bump within the expected range of ~20% (before tracing back near the IPO price on Friday). They raised close to $100B, and the stock closed at about a $2T valuation if you consider unvested shares. I have nothing novel to say on this very unobscure company, but a bizarre future awaits.
If it keeps running in an intelligence arms race, then does getting paid off on it require being the last company standing? And if it is, does that mean it’s even stronger than the government? Because if it still has to answer to the Air Force and the Supreme Court, then is the commander-in-chief also the CEO?
What if it slows down by redirecting capex to make products that knock off industries one at a time? That’s a FCF bonanza that will make the current valuation seem laughably small…for a hot second before it’s lapped by someone committed to the arms race who has to compete for capital with one less elephant.
I guess so much depends on whether the machine god they’re building will make room for us humans. Dario claims the god will be benevolent when it talks to people who can stop it, and he admits it won’t when it’s trying to sound important and worthy of unskeptical sums of capital. See Anthropic Is Not Living Up To Its Ideals.
This passage in particular encapsulates my mistrust:
Mealy-mouthed corporate platitudes might even be making the situation worse. “This technology could kill all of you, but I have to make it anyway” sounds hypocritical on its face, so the average person assumes that there must be some sort of ulterior motive: maybe marketing hype, maybe regulatory capture, maybe a liability shield; but the companies aren’t acting like they really actually believe their warnings. I’m sympathetic to the average person here. I think many people at the AI companies believe the dangers are real, but the companies as a whole sure aren’t acting like it.
This is quite the pretzel.
The general public sees a company throwing a humanity-sized noose over a tree branch and assumes they can’t be serious and that this spectacle is an unserious ransom note. Because how could it be anything else, right? You wouldn’t actually tie the noose if you thought it would work. Right?
Money Angle
For the bulk of our wealth, we are diversified so as to underwrite systematic (i.e., non-idiosyncratic) risk in exchange for the premium associated with that risk. The focus is on risk and our appetite for it, which then determines sizing.
The approach is mostly indifferent to expected returns, which are
a) far less predictable than vol and
b) less important since your compounded growth is far more sensitive to the square of vol.
It’s convenient that the focus on the easier thing is also a focus on the thing that matters more.
(And beyond that, the math of taxes is the next biggest lever since it’s high impact and more manipulable than returns, which you have zero control over)
If you weren’t already discouraged at the prospect of predicting returns, this observation suggests that even if you could predict earnings perfectly, you would still be missing 2/3 of the variation of returns.

But man, that’s a lot of dowdy throat-clearing.
We wanna talk about returns anyway, goddammit. Let’s do that in the context of the current market and learn some finance along the way.
In what if the “singularity” trade cracks? I used earnings yield + growth to bracket SPX returns with a floor around 5.2% and a ceiling around 19%. We’ll revisit that and compare the logic to a similar calculation from Elm Wealth.
Holding P/E constant (the tweet above reminds us just how fake that assumption would be, but we are in learning mode for a moment), a stock’s return comes from two sources:
-
The dividend. Payout ratio × earnings yield.
-
Growth from retained earnings. Whatever isn’t paid out gets reinvested. If it earns the company’s ROE, earnings grow at
g = (1 − payout) × ROE
Add them:
E[r] = payout × E/P + (1 − payout) × ROE
Two extremes bracket what happens at real companies:
-
Payout = 100% → E[r] = E/P. No reinvestment, so no growth. You collect the earnings yield.
-
Payout = 0% → E[r] = ROE. Everything compounds inside the business.
SPX metrics

The payout is inferred by what we can observe which is the actual dividend yield realized. The SPX dividend yield is about 1%, which represents about ~20% of earnings.
So the 80% of retained earnings is reinvested at the ROE and you back into a number that is much more aligned with annual SPX returns in recent years. The simple earnings yield is not accounting for the growth of reinvested earnings.
At 22% ROE: 0.80 × 22% = 17.6%
Add in the 1% dividend yield, and you arrive at E[r] ≈ 18.7%
But this is too aggressive.
Why?
The formula assumes every retained dollar earns the average ROE. A lot of SPX retention goes to buybacks, which only returns the earnings yield, not ROE.
If 100% of retention went to buybacks, then E[r]= 5.46%
This helps bracket a range of probable returns (again holding the multiple constant) from a low end of about 5.5% (all buybacks) to about 19% (all reinvested at average ROE). A key inference is that not only is ROE important, but also what portion of the 80% of retained earnings gets reinvested at that ROE.
[In what if the “singularity” trade cracks?, we see how Goldman expects AI capex to drag mega-cap ROE down about 7 points because marginal ROE on that spend is lower than the ROE of the first reinvested dollars.]
Just to inject a personal message. The first time I myself ever learned about decomposing earnings into a portion that is paid out and a retained portion that is reinvested was working through an Aswath Damodaran spreadsheet that he publishes when he updates his valuation model. (Sadly, this was likely only a decade ago that I learned this stuff). I recently got sucked into reading a Twitter thread where there was back and forth on this way of decomposing returns which I of course fed to AI to ELI5 and just found it all satisfyingly edifying that stock market returns are not so mysterious. We literally just have record ROE and lots of reinvestment. Of course, the part that feels most pressing today is the extent to which earnings and ROE numbers rely on circular booking for still unfilled orders for compute, chips, and general AI infra smorgasbord. From a personal education pov, I’m quite pleased to zoom in closer than a simple PEG model that says returns are earnings yield + growth but stops short of explicating how growth depends on retention and ROE.
It’s hard to place a section like this in the letter. It feels more advanced than what I might typically write in Money Angle, but it’s Crayola simple to anyone who was even an analyst and not worthy of being in the Masochists section. But then again, maybe the Masochists readers probably have melted derivatives brains, not a champagne corporate finance brain, so they don’t know this stuff…maybe that’s the target audience for this essay? I don’t know.
I do know that I enjoy learning this and just want to share, which is an instinct I’ve learned to trust.
Elm’s version: P-CAPE
At the same time I’m brushing up on this stuff, by serendipity or Baader-Meinhof, I noticed that the Elm Wealth squad uses a similar concept to adjust their market valuation for growth. They explain their method in Introducing P-CAPE.
To summarize it:
1. Start with Shiller’s CAPE: price ÷ the 10-year average of inflation-adjusted earnings. Flip it and you get CAEY, the cyclically adjusted earnings yield, which they treat as the expected real return if all earnings were paid out. (Notice how this maps to the floor scenario above where E[r] is only 5.46% except that’s in nominal terms.)
2. Adjust the 10-year lookback’s undercounting of earnings because it ignores what got retained (buybacks included) along the way. So for each of the last 10 years, they take the unpaid portion of earnings and compound it forward at that year’s CAEY, then average. Their example, 10 years back with a 60% payout, $10 of real earnings and a 6% CAEY:
60% × $10 + 40% × $10 × (1.06)¹⁰ ≈ $13.2
Price over that adjusted number is P-CAPE.
Its reciprocal, P-CAEY, is the return forecast.
They say that from 1890–2024, the P-CAPE earnings measure averaged 2% above the next year’s actual earnings, whereas plain CAPE ran 13% below. It explains 35% of the variance in 10-year forward real returns versus 24% for CAPE (33% vs 15% since 1950).
Comparing the two
The wedge in methods comes down to what a retained dollar earns.
We started by letting it earn ROE, but Elm lets it earn the earnings yield, claiming that the marginal investment yields the average return on equity already in place which just collapses to the full payout case. The earnings yield is the return. The payout adjustment goes into the earnings number instead of into a growth term stacked on top.
Their capital market assumptions as of 9/30/26 reflect this conservativeness, with US stocks implying a 2.92% real expected return (P-CAEY). Tacking on a breakeven inflation rate from Treasuries of about 2.20% brings the nominal projection to ~5.12%, a touch less than our 5.46% floor.
The gap between the floor of about 5% and the ceiling of about 19% is 14 return points. If you took the midpoint and say E[R] is 12% give or take 7 points. That sounds low compared to a 1-year SPX straddle, which at 18% vol would give or take about 14 points* around a mean of a 4% RFR.
But don’t forget…most of the volatility is not in the earnings but the change in the multiple! If God decreed that earnings volatility only accounted for 1/3 of total volatility, with the rest attributed to changes in multiples as Brett’s tweet alluded, then we’d expect the SPX should be closer to a 21-point straddle or about 50% more than it’s priced. The market’s best guess for the price of the straddle (indeed ~14 points) is worth accepting, so I think that means our range of E[R] due to earnings only of 5 to 19% is just excessively wide.
*straddle is 80% of a standard deviation or vol
Money Angle For Masochists
Using the t-Statistic to Validate Your Trading Results | 4 min read
Easy statistics for A/B testing and hamsters | 7 min read
Some stats education.
The articles are two approaches to a common question: how far is this result from nothing, relative to the noise you’d expect?
-
Kaabar divides a strategy’s average return by its standard error (0.05% a day, 1% daily vol, 1,000 days gives t ≈ 1.58).
-
Cohen boils the chi-square test for a 50/50 split down to D² > N, where N is total clicks and D is half the gap. That’s an absolute z-score above 2, a close cousin of t ≈ 2.
The Cohen trick is pretty neat so I’ll quote:
When you start applying the formula to real-world examples, you’ll notice that when N is small, it is difficult—or even impossible—to be statistically significant. For example, say you’ve got one ad with 6 clicks and the other with 1. That’s N=7;D=2.5;D² =6.25. So the test is still inconclusive, even though A is beating B six-to-one. Trust the math here—with only a few data points, you really don’t know anything yet.
The smaller the N, the bigger the difference needs to be, to be detectable… So for example, if N=50, as it was in our “day one A/B test” example, the winning variant needs to be almost double the number of clicks as the losing variant in order to have a detectable difference, e.g. a conversion rate of 10% versus 5%. This is a huge difference; it’s great if you find something so dramatically better, but this is rare, and therefore you can almost never find a significant A/B test given only 50 clicks to analyze.
When N=100, the winner needs to be at least 50% higher than the loser (which it was by the second day in our example). It takes N=1800 to detect the case where the winner is only 10% larger than the loser.
And this is bad news for A/B tests, because often one variant isn’t better than the other by more than 10%, e.g. a “2.4% conversion rate” versus a “2.2% conversion rate.” What does this mean, especially if you don’t have large N? It means you need to be seeking big differences, not subtle ones. Test wildly different designs, rather than tweaks. Tweaks can only be tested when N is enormous.
[The articles look like they disagree because Cohen’s footnote says t-tests are the wrong tool for A/B tests. but he means counts of outcomes. Returns are measurements, so the t-stat is a fine tool for Kaabar’s question.]
Stay groovy
☮️
