How to tell if your investment thesis is broken (or if you just had a bad quarter)

Every piece of writing on when to sell a stock arrives at the same sentence and stops before the part people most need to know: sell when the thesis is broken. It tells you to watch for fundamentals breaking down, to be intellectually honest, to distinguish a broken thesis from a temporary setback. None of it tells you exactly how.

That is because the how is the hard part. A real process has to name the point in time at which you were wrong, and naming it is what everybody, writers and investors both, would rather avoid. That’s why the gap exists.

Here is the test. A thesis is broken when a driver you wrote down has stopped being true on the terms you set in advance, and the evidence comes from a document rather than from the price or mood. Everything else is a bad quarter, a bad month, or a bad vibe.

If you have not written a thesis down for a stock, you cannot run this process, because you are not evaluating a position. You are arguing with yourself about one, and you will lose, because the version of you holding a losing position is not a neutral party.

If you need the structure first, start with what an investment thesis is and how to write one. This post assumes you have one.

What does “broken” actually mean?

A thesis is a set of claims. Each claim rests on one or more drivers, which are the mechanisms that have to hold up to stay true. Each claim has an invalidator, which is the specific observation that would prove the claim wrong, with no room for interpretation.

So at any moment a claim sits somewhere on a spectrum:

• Confirmed. A filing or a print directly evidences the driver.
• Supported. Evidence points toward the claim without settling it.
• Holding. Neutral. Nothing has touched the claim, or nothing that touched it moved its status. This is where most claims sit most of the time, and it is not a failure state. It is the default.
• Drifting. Evidence has begun to point away, below the threshold you set.
• Challenged. Something material has landed against the claim and has not resolved.
• Broken. The invalidator tripped, on the terms you wrote when you were calm.

Notice the one thing that is missing from the list, and it is the thing people look at most. The price. A position can be down 40% with every claim holding, and up 40% with every claim broken. The market is voting on a different question than the one your thesis asks.

Most investors cannot answer whether their thesis is broken because they are looking at the only piece of evidence guaranteed not to contain the answer.

The mistake many make: outcomes are not drivers

This is the distinction that does the most work, and it is the one the standard advice collapses.

Let’s take the usual list of warning signs. Declining margins. Shrinking revenue. Rising debt. Lost market share. Every one of those is an outcome. By the time an outcome moves, the cause has usually been in motion for months or quarters and the price has already reflected it. You are trying to catch the cause, not read the receipt.

A driver is the mechanism underneath the outcome. You do not own margin expansion. You own whatever is causing it, and that is the thing that can quietly stop.

Here is the same distinction across a few common situations.

Revenue missed consensus. Consensus is not your claim unless you wrote it down as your claim. If your driver was “subscription revenue grows above 8%,” and subscription revenue grew 9% while total revenue missed on a hardware line you never mentioned, nothing in your thesis moved.

The stock fell 12% on the print. The market repriced. Your claims are unaffected until you can point to the line item that touched one. Most of the time you cannot, because the move was multiple compression, positioning, or a sector rotation that arrived on your ticker.

A competitor launched something. This is more ambiguous, and it is where most people either panic or wave it off. The test: does the launch change the mechanism, or does it change sentiment about the mechanism? A competitor announcing a product is sentiment. A competitor disclosing signed capacity or a named customer win is mechanism.

Management changed the language. Underrated, and often the earliest real signal. If last quarter’s MD&A said demand exceeds supply and this quarter says demand remains healthy, something moved before any number did.

The three ways a thesis actually fails

Not all breaks are the same, and treating them the same is how people either hold too long or sell too early.

1. The driver was invalidated

The cleanest case. You wrote “gross margin below 70% for two consecutive quarters” and it happened. The mechanism you were counting on stopped working.

This is the one your written invalidator is built for. Act on the rule you wrote, not the feeling you have now.

2. The driver held, but slower than you assumed

This is common, and the hardest to handle honestly. The mechanism works. Your timeline did not.

This is not a broken thesis. It is a repriced one. The claim is intact, but the return you were underwriting has thinned because the same outcome now arrives later, and capital tied up for eight quarters instead of four is a different trade.

The honest move is to re-underwrite from today, not to quietly extend the horizon and keep the original story. Ask: at today’s price, with what I now know about the pace, would I open this position? If no, the thesis is fine but the position is not.

3. The claim was never specific enough to fail

The most common, and the least discussed. You did not write a thesis. You wrote a description of a company you liked, in language that could survive any outcome.

“Strong management.” “Great product.” “Well positioned for AI.” None of these can break, which is exactly why they feel comfortable and exactly why they are worthless.

You cannot fix this retroactively without lying to yourself, because you now know what happened. What you can do is rewrite the claim as of today, in falsifiable terms, and start the clock from here. Do not backdate it. Backdating a claim to match what already occurred is how you learn nothing.

What can invalidate a claim, and what can’t

Not all evidence has the same weight. Roughly in order:

Primary and binding. 10-K and 10-Q line items, segment disclosure, and the language in MD&A. An 8-K for material events. These are the documents a company signs its name to, and they are where a driver either shows up or does not.

Primary and directional. The earnings call. Prepared remarks are a marketing document, so read them for changes in language rather than content. The Q&A is more useful, because analysts ask the questions management did not plan to answer, and the shape of a non-answer is information.

Secondary. Reported news with a named source, a competitor’s own filings, regulatory filings and decisions, supply chain disclosure from a customer or supplier. These can touch a claim, and occasionally settle one.

Not evidence. Price. Analyst ratings and target changes. Headlines that restate the price move as a cause. Aggregator posts. Anything phrased as “investors are worried about.”

That last category is where nearly all of the volume is, and where nearly all of the anxiety comes from. A mid-cap company generates perhaps three or four documents a quarter that could genuinely move a driver, against several hundred headlines. The ratio is the whole problem.

On the NVDA thesis below, 593 events were screened. Seven claims. Most of those events touched none of them.

What each status looks like in practice

Here is a live thesis on NVDA, published August 25, 2026. The claims were written first. Everything under them is what the engine found.

The thesis: Long NVDA because AI compute demand is still outrunning supply: data center revenue ran $75B last quarter, up 92% y/y, and next quarter is guided to ~$91B with zero China compute assumed. Hyperscaler capex, tracking toward $700B this year, is fuel, and CUDA lock-in means Nvidia captures most of that spend at ~75% gross margins.

That sentence breaks into seven claims. As of August 26 they sit in four different places.

Supported. AI compute demand continues to exceed available supply for Nvidia data center products. Capacity still looks tight, with buildouts and backlog growth pointing to demand absorbing systems faster than supply. One early sign cuts against it: efficiency gains can trim compute needs. 31 alerts.

Supported. Hyperscaler capital spending remains elevated. Customer spending and supplier demand both hold up. One minor warning, again on model efficiency. 69 alerts.

Holding. Data center revenue continues growing at a strong year-over-year pace. Nothing has moved it. The next confirmation is the reported number. 7 alerts.

Holding. Nvidia meets or exceeds next-quarter data center guidance despite assuming no China compute. No direct read yet. Stands pending results. 7 alerts.

Challenged. CUDA remains a strong customer lock-in factor. The ecosystem still shows stickiness, but live deployments on competing platforms show the lock-in is not as exclusive as the claim assumed. 42 alerts.

Drifting. Nvidia continues capturing most hyperscaler AI compute spending. Still winning major funded projects, but part of hyperscaler and model-provider spend is now landing on internal chips and rival systems. 50 alerts.

Drifting. Nvidia sustains gross margins around 75%. Still plausible, but discounting and margin pressure elsewhere in AI infrastructure are early signs it is getting harder. 36 alerts.

Two things worth pulling out.

Nothing is broken, and the thesis is still not simply fine. The overall read is supported while two claims drift and one is challenged. That is what a real position looks like most of the time. The headline status is a summary, not the answer. The answer is which claim moved. Here, it is the two the original sentence leaned on hardest: capturing most of the spend, and holding 75% margins.

A claim that starts pointing away is not a claim that failed. Neither of those has broken. The correct move on a drifting claim is to write down what would settle it and wait for the next print. That is the hardest thing to do, and it is why people either dismiss it or sell.

See the full thesis

The two ways to get this wrong

There are failure modes on both sides, and they are mirror images.

Thesis drift. You quietly rewrite the claims so they match what happened. The horizon extends. The invalidator softens. The story mutates until it can never be wrong. This is the more common failure and it is almost invisible from the inside. Each individual edit feels reasonable, and if the thesis only ever lived in your head there was never a version to edit against. The story just becomes what you now remember it.

The defense is mechanical: keep the original text, with dates, and never edit in place. Add a revision with a reason. If you cannot articulate why you are changing a claim without referencing the price, you are drifting.

Premature selling. You sell on the first challenged event because the position is uncomfortable. The mechanism was intact and slow, and you converted a timing problem into a realized loss.

The defense is the same document. If the invalidator did not trip, and you cannot name the driver that failed, you are selling a feeling.

Both failures come from the same root: making the decision in the moment rather than checking it against a decision you already made.

The ten minute check

Once a quarter, per position:

1. Reread the original claims. The original text, not your memory of it.
2. For each driver, find the number or the disclosure in the most recent filing. Not the summary, the filing.
3. Mark each claim intact, challenged, or broken, using the invalidator as written.
4. For anything challenged, write one line on what would settle it, and by when.
5. For anything broken, act on the rule you wrote.
6. If you changed a claim, write why, below the original, with a date.

That is it. Roughly ten to twenty minutes per position per quarter. Which is fine at three positions and very hard once you get to fifteen, and that constraint is the honest reason most people stop doing it after two quarters.

Where Thesis helps

This is the gap Thesis was built to close. You write the claims and Thesis reads the filings, the earnings transcripts and the news, and speaks only when something touches one of them.

Concretely, a margin miss like the one above reaches you looking like this.

The alert reports what the event touched. It does not tell you the claim is wrong, or what to do about the position. That call is yours, and it should be, because you are the one who wrote the claim and you are the one holding the risk.

And on a day the stock drops 18% with no company news: nothing. Silence is a feature. Most days nothing has happened to your thesis, and you should be able to trust that.

Frequently asked questions