Case study

I tested whether a podcast can pick stocks.

Turns out: no. I fed every All-In episode since ChatGPT launched into Claude, scored what the guys actually backed, then checked it against what those stocks did next. The picks lost to their own theme by 5.9% a quarter. That killed the version of this project I wanted to be right. Here's what replaced it.

The live book, since inception

The framework, in plain terms

The one-sentence version

The podcast is good at spotting a durable theme and bad at everything else. So buy the theme through funds, hold it, and use the show only to watch for the theme breaking.

Why it's built that way

176 episodes became 176 hypothetical portfolios and 175 weekly rotations, tested across four market regimes. Four questions came back with four answers.

QuestionAnswer
Can they pick stocks week to week? No. t = 0.72, 50.9% hit rate
Can they time the market? No. r = -0.06 with forward returns
Are their high-conviction picks better? Worse. t = -2.26
Was the theme they kept repeating right? Yes. +385% against the S&P's +85%

The third row is the load-bearing one. It's the only statistically significant result in the project, and it points the wrong way. The names they backed hardest, GOOGL in 113 of 175 episodes, then MSFT, META and TSLA, score high because they're newsworthy, not because they're good ideas. Conviction was measuring newsiness.

So: harvest the theme, ignore the picks.

The thesis

AI compute is supply-constrained, and the constraint moves

Demand outruns capacity. First chips, then memory, then power, then the physical datacenter buildout. Whichever layer is scarcest captures the pricing power.

This is a claim about mechanism, fab lead times and grid queues and capex commitments, not a bet that past returns continue. That distinction is the whole point: it makes the thesis falsifiable.

RegimeSMHS&P 500
2023 ignition+56.9%+17.7%
2024 bull+40.5%+24.1%
2025+47.3%+15.7%
2026 run and crash+46.5%+8.9%

Never a losing regime. Including the one with a 36% drawdown in it.

What's in the book

Three sleeves

Theme, the proven core

Two funds rather than one. SMH is cap-weighted and runs about 22% NVDA. AIS is supply-chain weighted, caps NVDA near 3%, and holds the memory and power names. Together they cut look-through NVDA from 21.7% to 13.6%.

Migration

Power and physical buildout: the layer SMH structurally cannot hold, and the one the show has been loudest about. If the bottleneck moves there, this is what catches it.

Experiment

Deliberately holds the signal proven negative, sized small. A live forward test: if it lags, t = -2.26 is confirmed out of sample. A dead hypothesis deserves a verdict, not a footnote.

Always 100% invested. Their average 22% cash cost 36 points of return in the backtest, so there is no default cash position, ever.

The rules

Four of them

1. Don't pick stocks

Fund managers choose the constituents, by published rules, written by people who didn't know the outcome. This is the direct fix for the t = -2.26 problem.

2. Rebalance quarterly, never weekly

Weekly alpha was zero. Weekly tinkering ran -0.56% a week against the very sector it was picking from.

3. Stay fully invested

No cash, no leverage, no shorts, no timing. Bearish views get expressed by absence.

4. Don't touch the parameters

The one that actually protects the other three, and the one aimed squarely at me.

Rule 4 is the important one

A 108-configuration parameter sweep produced an in-sample versus out-of-sample correlation of -0.21. Negative. The best-fitted setup lost to the S&P out of sample.

Which produced the governing principle: backtests are used to reject constructions, never to select them.

Every positive backtest result in this project either reversed or turned out contaminated. That includes a +718% basket I nearly pitched before catching that I'd picked its constituents already knowing which names won, and a +734% sector filter that returned +53% out of sample against the theme's +109%. Every negative result held.

So the framework changes for exactly three reasons: a genuine bug, a structural break in the thesis, or three years of forward data. Explicitly not "it had a bad quarter." Three of my own conclusions have already reversed under more data, which is why the guardrail exists.

What the podcast still does

One job

It no longer picks anything. Each week it's watched for exactly one thing: has the bottleneck moved, or has the thesis broken?

The thesis dies if hyperscaler capex guidance turns down, or if supply catches up across chips and memory and power at once with no new scarce layer appearing. Monitored weekly, flagged for a human, never auto-traded. Note what isn't on that list: a bad quarter.

The live book

Where it stands

Paper-tracked, $100,000 notional. No real money moves, ever.

Holdings
TickerJob WeightValueReturn

Risk and benchmark

What the allocation actually does

Backtest: target weights over history
Return and risk
TotalCAGR VolatilityMax drawdownSharpe
Relative to each benchmark
BenchmarkBetaAlpha /yr Tracking error Info ratioUp captureDown capture

Reading it honestly

Against SPY, this is just beta

An alpha of a year against the S&P looks incredible and means basically nothing. Beta to SPY is , R² is . It's a semiconductor book measured over a semiconductor bull run. SMH is the benchmark that matters, which is why the framework names it.

Against SMH, it's a draw

The book trailed SMH on raw return, but at lower volatility and with a shallower max drawdown. Sharpe landed at against SMH's . The diversification roughly paid for itself. That's the whole claim.

The information ratio is negative

At , every unit of tracking error I took against SMH destroyed value instead of adding it. That's the honest verdict on the active part of this thing, and it goes on the page right next to everything else.

Down capture contradicts a stated goal

The framework says success looks like shallower drawdowns than SMH. Max drawdown agrees: against . Monthly down capture disagrees, at . Both are true. The gap is the AIS sleeve amplifying monthly declines while the peak-to-trough path still came out shallower.

The window is short and flattering

months, bounded by the newest holding, covering one unusually good stretch for one asset class. A CAGR is a fact about that window, not a forecast. None of this has seen a real bear market in semis.

It is a backtest, so it can only reject

This runs today's target weights over the history those holdings actually had. It is not the live track record, which is days old. Per the anti-tuning rule, numbers like these get to kill a construction. They never get to justify one.

Holding correlations, daily returns

The between SMH and AIS is the concentration the framework already flags. 70% of the book sits in two vehicles that move together.

Scoring myself

What "working" looks like

It's paper-traded, so success isn't "it went up." Four things are, and none of them is the return.

Stays fully invested

Tracking the theme, no default cash.

Shallower drawdowns than SMH

The diversification earning its keep, rather than just costing return.

The rules run without me

No discretionary overrides. Hardest one on the list.

The experiment returns a verdict

Either way, and I publish it either way.

The honest caveat

This does not beat just buying SMH

Over the window measured on this page the book returned against SMH's , so it trails by . Construction testing over a longer window put the same gap nearer 6 points. Either way the direction is the same and it is not in the book's favour.

What that buys is migration insurance and a shallower hole: max drawdown against , Sharpe against . That Sharpe gap is well inside noise and should not be read as skill.

If maximum return were the only goal, the evidence says raise the theme weight, not lower it. The reason to hold the diversified version is that it's the one you don't abandon during a 36% drawdown, and the one that survives if the bottleneck moves somewhere SMH structurally can't follow.

The sequel

Same idea, pointed at something bigger

Grading a podcast was the small version. The AI Stack is an agent doing it to itself every weekday: it maps a layer of the AI buildout, makes a call with a resolution date, and publishes what the reviewers killed. It opened by generating fourteen ideas and killing all fourteen.

Hypothetical paper portfolio built from podcast commentary and tracked for research. Not investment advice, not a recommendation to buy or sell anything. I'm not a licensed financial advisor.