Do Rolex prices predict market crashes? I ran a study in Claude Code...
By Felipe SinisterraJuly 29, 20269 min read
I was on X this week and saw this.
Nobody needs a Rolex. It is the first thing people of means sell when they want cash without touching anything that actually hurts.
Behind it is a hedge fund manager who has been closed to new money since the early 2000s, so this is not somebody guessing.
It is also the kind of claim that sounds right, gets repeated for years, and never actually gets checked.
The great thing about AI is that you can check it now. Even on something like watch prices, which used to sit inside a dealer network where nobody outside the trade could get at it.
So that is what I did this week. I backtested the theory against seven years of daily watch prices to find out whether it is true.
(Of course using AI, and I’ll show you how)
Here’s the plan:
Name what the claim is actually about.
Check what the data is capable of showing.
Write the failure line before loading anything. S
core every firing against the price chart.
Run the whole test backwards.
The Prompt
I ran this in Claude Code.
Watch index vs. S&P 500 and bitcoin since 2019
The five moves are the technique. The watch market is just what I pointed them at.
Name the observable. The post was about supply, meaning listings flooding the market. The only public data is a price index, and those two are not the same thing. Most inherited claims die right here, in the gap between what somebody asserts and what you can actually measure.
Check what the series can show. How a number is built decides how fast it is allowed to move. The watch index behaves like a multi-week average, smooth enough that it cannot show a fast reset even if one happened, and most of my history predates the day the index went live, so those years were filled in after the fact rather than published at the time. There is a thirty second version of this check that kills most claims for free: if the series that supposedly leads is published more slowly or more smoothly than the series it supposedly leads, stop before you open a spreadsheet.
Write the failure line first. One sentence with numbers in it, covering which direction, how large, and inside what window. Then ask whether your sample could detect a move that size if it were real. Decide what counts as failure after you have seen the first chart and every ambiguous reading gets scored as a point for the claim.
Score every firing, and beat the price chart. Count the times the signal fired and nothing followed, and report how much of the time it sat switched on. Then make it beat the free alternative already on your desk, which is the price trend of the thing you would be trading. A rule that is on most of the time will eventually sit in front of every selloff, and that coverage reads like foresight when you only count the hits.
Run it backwards. Reverse the arrow and run the identical test, then look for a listed instrument carrying the same exposure on a faster clock. This is the move that turns a dead claim into something you can actually put on.
What you are buying here is coverage. Every episode in the record gets scored the same way, including the ones nobody quotes, and running the test backwards costs the same as running it forwards.
The prompt / paste into Claude Code
I want to test a market claim I picked up from someone else, and I need the test built so that it is able to come back no. The claim is: [PASTE THE CLAIM EXACTLY AS IT WAS WRITTEN, PLUS THE DATES THE PERSON SAYS IT APPLIES TO].
Your first response must contain the plan below and nothing else. Numbers you choose in advance as thresholds are expected. Numbers that describe what actually happened are not, so do not pull, compute or estimate any figure from data until I have approved the plan and given you a file. If you cannot compute something from a file I supply, write "not computed" and tell me what would make it computable.
1.
Name the observable. Tell me what the claim is actually about in the claimant's own terms, then tell me whether the series you can get measures that same thing or only something correlated with it. If those are different, say so plainly, because that gap often settles the question on its own.
2.
Check what the data can show. For every series you propose, give me how it is built, how often it publishes, its reporting lag, and any smoothing or rolling average inside it. Then answer one question directly: can a series built this way answer a timing question at all? If the series that supposedly leads is slower or smoother than the series it supposedly leads, tell me to stop there.
3.
Rank your sourcing. Name the ideal series first, then a free downloadable substitute, then a listed instrument that carries the same exposure. If a provider publishes no download, check whether the chart on the page loads from a public data endpoint before you tell me a series is unavailable. Tell me which of the three you ended up on and name the source.
4.
Write the failure line. One sentence stating the result that would end this claim, with numbers in it: direction, size, and the window it has to happen in. Then tell me whether the sample can detect a move that size if it is real, and if it cannot, say so before we run anything.
5.
Set the scoring rules. Define the test episodes mechanically across the history the data allows, give me that definition and the exact fields of the episode list now, and print the full list before you compute a single result. Every firing gets scored, including the ones followed by nothing, and I want the share of time the signal sits switched on. Grade the rule against the same rule run on the price trend of the asset I would be trading.
6.
Plan the reverse test. The identical test with the arrow pointed the other way, including one listed, liquid instrument as a candidate leading series.
Keep the plan to three series or fewer so this runs in a single session. When you report results later, include the tests that found nothing, and flag any departure from the approved plan at the top of your answer.
Swap the episode threshold in step 5 and the candidate series in step 3 for whatever your own claim runs on. If your claim is about flows, tell it to grade the signal against the trend in that flow series.
Thirty minutes buys you the plan and a verdict on whether the data can answer the question at all. That verdict is where most inherited claims die.
The mechanism the claim depends on actually fired this year. What it was supposed to cause never showed up.
Supply really did flood. Listings hit record levels in April and May, and search interest in selling a watch reached its highest reading in the seven years of data I have, with all three sell-side search terms maxing out in May.
Sellside search interest against the watch price index. Selling intent spiked in sprint 2026 while prices went up.
The price index is up 3.4 percent since the end of March, and it set that 2026 high on July 13.
So the flood happened, and the reset did not.
Across the nine equity drawdowns since 2019, watch prices rose during five of them and finished within one percent of where they started in two more. Only two episodes saw watches fall more than a percent, and the worst of those was 2.8 percent.
The watch index through every equity drawdown since 2019
The 2021 top is the cleanest picture of the ordering. Bitcoin peaked on November 8, the Nasdaq on December 27, the S&P on January 3, and the watch index on March 20, a full 83 days after the Nasdaq.
It gained another 18.6 percent after the S&P had already peaked.
Bitcoin, the Nasdaq, the S&P and the watch index (indexed to same starting point)
Through COVID the watch index fell 2.6 percent while the S&P fell 34 percent.
Graded as a signal it gets worse. Watch momentum was negative in 165 of 368 weeks, and only 12 percent of those weeks were followed by a correction inside a quarter, against a base rate of 22 percent. One version of the signal sat switched on for 216 straight weeks.
Trading it did not help either. No watch rule beat the same rule run on the asset’s own price by a margin you could call real, and the one that came closest was the version that traded least.
Every week the watch stress signal was switched on, against the equity corrections that actually started
Then I ran it backwards, and it worked. Equity moves predict watch prices about a quarter later, and Watches of Switzerland, the listed retailer, prices the state of the watch market roughly 11 weeks before the physical market gets there.
The information in this claim is real. It runs the opposite way from the version going around, and it is already sitting on an exchange.
Three limits worth stating plainly. I tested a price index and not listing data, so the supply half of the claim is only settled by what prices went on to do.
The index itself only went live in late 2023, so roughly two thirds of my history was filled in afterwards and nobody could have traded it at the time. Read this as a study of the record, not a strategy you could have run.
And last year’s Swiss tariffs lifted watch prices for reasons that have nothing to do with anyone’s risk appetite.
The five moves work the same way on a sell-side thematic note, an expert network call, or the rule of thumb a colleague has been repeating since 2011. That is where most of these claims actually reach you.
Personal
Speaking of watches, I've been getting into fashion lately. Saw an open source project blow up on X this week.
The code is free for anyone to take, so I've been rebuilding it into the version I actually want. Further down the rabbit hole than I planned.
There's a ton of software sitting out there like that. If you want some thoughts on how to take one and make it yours, reply and I'll send you a few tips.
Enjoyed this article?
Get more AI for finance content delivered to your inbox.
