How to research the next big trade

After looking at the rally in AI-linked drug-discovery suppliers, I wanted to understand what an investor could have seen a year earlier.

I want to reconstruct the information available at the time, including what the public clues supported and where an investor still had to make a leap of faith.

That gives me somewhere to start on the next opportunity: take a change in the world, work out which businesses it affects, and make a forecast that can be compared with their prices.

Start with what was knowable

Several parts of the AI-biology thesis were visible by October 2025.

In January, NVIDIA announced collaborations with Illumina, IQVIA and Arc Institute. That identified organizations working on biological models, data and clinical workflows.

In February, CZI launched its Billion Cells Project with 10x Genomics and Ultima Genomics. The announcement named technology intended to generate the data.

In August, GenScript reported 52% growth in its protein business and pointed to AI drug-discovery demand.

The announcements covered a collaboration and a specified data-generation project; GenScript's results added reported commercial growth. Together, they gave investors a reason to investigate biological AI as a source of laboratory spending.

None of them established which security would perform best or why a particular investor bought it.

What interests me is how the clues connect. The model announcement described a technical possibility. The data project named a physical input, and supplier results began showing whether customers were paying for the work.

A primary-source announcement records what its issuer said or committed. I would not treat it as independent proof of the whole thesis. A sponsored benchmark, a customer contract and an audited cash-flow statement answer different questions.

Publication dates, reporting periods and expected completion dates need to stay separate. A full-year report released in 2026 cannot show what an investor knew in October 2025. Even an apparently historical URL can lead to a page containing later updates.

For each clue, I would save the original document and its publication date, then record the claim I used and what I inferred from it. Several tweets repeating one company announcement still amount to one underlying source.

Keep the candidates that disappointed

Once we know the winners, it is easy to assemble the announcements that make their success look inevitable. The reconstruction needs the plausible alternatives that disappointed too.

Recursion is one of those alternatives.

For example, Recursion's August 2025 results included a Sanofi milestone, clinical programs and an AI platform. The same release reported $208.4 million of operating cash outflow for the first half. An investor examining AI drug discovery then had reasons to investigate the company and reasons to question its financing needs.

Could a research process have distinguished among these opportunities using the evidence and prices available then? Looking only at today's favored suppliers cannot answer that.

Knowing the outcome influences even a carefully reconstructed list. A dated watchlist created before the outcomes offers a stronger test.

I would create a historical candidate list using a rule established before examining returns: include publicly listed suppliers or platforms explicitly named in the dated evidence, then investigate their material financial exposure. The sources above lead to companies such as Illumina, IQVIA, 10x, GenScript and Recursion; a synthesis supplier such as Twist needs its own contemporaneous operating evidence.

For each candidate I would record why it qualified, what would have disqualified it and the valuation an investor could observe then. Starting with today's winners and searching backward for supporting announcements rewards a convincing story.

Broad equities and biotechnology provide benchmarks for the candidates. A stock can rise and still underperform its relevant market. These benchmarks describe relative performance, but they do not automatically adjust for currency, size, interest-rate sensitivity or other risks.

The reconstruction can tell us whether the thesis was discoverable. Showing that the method helps select investments requires more: a prospectively recorded process, repeated opportunities and results after costs.

Follow the change into a customer's budget

The proposed investment depends on a sequence of changes:

A capability improves → customers change behavior → spending changes → a business captures profit → shareholders receive value.

Each step needs evidence. Progress at one does not establish the next.

Suppose a model makes molecular design cheaper. A pharmaceutical company could launch more projects. It could also keep the same projects and reduce research costs. It could buy more external testing, or invest in its own laboratory.

A model-quality announcement gives us no way to choose among those spending outcomes.

Papers and benchmarks help assess what has become possible. Customers' plans indicate whether they intend to adopt it. Suppliers' orders and repeat purchases show whether money is moving. Their financial statements show how much becomes profit and cash flow.

A customer paying repeatedly would carry more weight for me than another broad partnership announcement. Even then, I would check whether the spending is material to the supplier's total business.

Look for scarcity and for substitution

Calling every supplier a beneficiary misses the possibility of substitution. An improvement can increase demand for one input while making another less necessary.

In AI biology, better models may expand research while reducing experiments per project. Isomorphic Labs explicitly describes using computational search to narrow physical testing. Its claims need independent evaluation, but they identify a competing mechanism.

A useful simplified model is:

External laboratory spending = programs × experiments per program × outsourced share × price per experiment.

Which factors rise, and which fall? Even if demand outgrows supply, the shortage will only support durable profit if competitors cannot rapidly add equivalent capacity and customers cannot readily substitute.

For any technology that changes a workflow, I would ask which inputs it complements and which it replaces. Then I would identify the customer making that choice.

Build a serious competing explanation

I want a competing explanation for the encouraging observations, including those that seem to confirm my view.

A laboratory supplier might report higher revenue because conventional biotechnology funding recovered, because it gained market share, because it acquired a competitor, or because AI customers bought more services. Those explanations imply different durability and valuation.

Whole-company sales growth tells us less about causation than a customer identifying an AI program and placing a measurable order. A fixed cohort of comparable customers would help separate new demand from changes in customer mix.

Failures and abandoned work belong in the comparison too. If a platform reports a higher success rate after rejecting almost every candidate, I want to know how many useful candidates it discarded and what the selection process cost.

For AI biology, I would test four broad possibilities:

These outcomes could coexist across modalities. I would look for evidence that separates them; the same positive headline should not make me equally confident in all four.

Use Popper to specify how the idea could fail

I use Karl Popper's idea of falsifiability to ask what evidence would count against a claim.

"AI will transform biology" can survive almost any earnings report, which makes it hard to use in an investment decision.

I could test a narrower claim: "A specified supplier will show at least 25% growth in a relevant business over the next year, while maintaining its margin."

It has a date, a baseline and a possible failure. I would still need to check for acquisitions and changes in segment definitions. Either could create reported growth without confirming the mechanism, as could stronger conventional demand.

If the observation fails, it challenges the link I specified, without refuting every possible version of AI medicine. I would keep that failure on record instead of quietly replacing the original claim with one that fits the result.

Forecast a year ahead with three clocks

Technology, business results and market prices move on different schedules.

A capability can improve before customers adopt it. Revenue can grow before a business generates cash. Investors can anticipate both developments years before they occur.

For a one-year forecast, I would start with work already underway: funded projects, installations, customer qualification and upcoming results. Then I would distinguish what could become observable within a year from outcomes that require a much longer development cycle.

Twist reported $56.6 million of quarterly DNA Synthesis and Protein Solutions revenue for June 2026. I could use that quarter as the baseline for a forecast with fixed definitions:

Before relying on the forecast, I would assign a probability and record the reasoning. Putting a number on a judgment does not make it scientific.

Alongside the probability, I would record what could change it: repeat orders, customer concentration, price reductions, or evidence that customers are doing more work internally. Updates should preserve the original forecast so we can see how our reasoning evolved.

A provisional forecast sheet for the next year

The revenue target covers only a small part of the thesis. I also want to track adoption, supplier economics and whether the activity reaches manufacturing.

The following forecasts use an evidence deadline of October 6, 2027. These are subjective starting probabilities. I have not fitted them to historical frequencies, earnings consensus or options markets. They will be useful only if I record the reasoning, update them honestly and evaluate the results.

The 10x threshold is 15% above its reported $149.4 million comparable June 2026 revenue. That quarter's growth was 3%, so this demands an acceleration. Its August results provide the baseline. Twist's August results provide its margin and operating-income baselines.

The first three forecasts concern financial results. GAAP means generally accepted accounting principles; the operating-income test retains costs a company might exclude from an adjusted measure. A positive result still would not show that AI caused it. I would check acquisitions, settlement payments and segment changes. If the accounts cannot be reconciled to a comparable business, I would mark the forecast unresolved and keep the missing result visible.

The last three forecasts concern public evidence. For the manufacturing event, I would require an announcement identifying a purchase commitment, relevant supplier work and a therapeutic program connected to AI discovery or design. A generic partnership would not qualify. For the data event, I would exclude simulated cells, unperturbed controls and repackaging of previously released data.

If no qualifying disclosure appears, I would score that specified public event false. Private activity may still have occurred; the forecast measures what becomes public.

A broad funding change could affect several events together, so the forecasts are correlated. Adding their probabilities would be meaningless. Six related forecasts also cannot establish general forecasting ability.

Compare the forecast with the price

I could forecast the business correctly and still lose money on the stock.

The outcome I expect needs to be compared with the one that would justify the current price. That means estimating future margins and capital needs, allowing for dilution, and deciding what return would compensate for the risk.

A scenario that works only with rapid growth, high margins and a generous future valuation is fragile even if each assumption sounds individually plausible.

When reconstructing a past investment, I would use the share count, balance sheet and guidance available at the time. Applying today's financial information to last year's stock price would smuggle the answer into the exercise.

For an investment today, I would record several plausible outcomes and identify where I think investors may have misjudged the business. Consensus forecasts help with that comparison, but they cannot fully describe the expectations embedded in a price.

Reverse valuation gives me a way to quantify the disagreement: what growth and profitability would the entry price require to earn an acceptable return?

Suppose, with entirely invented inputs, a business trades at 20 times sales. We assume a five-year horizon, 2% annual share-count growth, an 8-times-sales exit valuation and a required annual return of 12%. The revenue growth required is approximately 37.2% per year.

Each assumption deserves investigation. Share-count growth reflects dilution: the business may grow faster than the ownership claim held by each investor. A high exit multiple assumes future buyers remain willing to pay generously for the business. Sales growth without a plausible cash margin cannot justify that multiple.

I do not need to agree on fair value to identify the assumptions a purchase relies on. I can then check whether customer budgets, capacity and economics support them.

The bear case needs its own forecast. Adoption could take longer, prices could fall, or a concentrated customer base could expose the supplier to a pullback. Weak cash conversion or a financing need could reduce the value of being right eventually.

The choice between shares and options adds a timing judgment. Shares have no contractual expiration date. A purchased call can expire worthless before a sound industrial thesis produces returns. A call spread changes the cost and payoff but caps upside. For an options trade, I would need a dated expectation change and an analysis of the actual premium; the long-term narrative cannot select the contract.

Keep a record that can teach us something

I want each serious idea to leave a short record:

• The change in the world and the customer behavior it should cause.
• The company expected to benefit and the mechanism connecting demand to profit.
• A dated forecast, probability and explicit failure conditions.
• The valuation assumptions and plausible downside.
• Subsequent evidence, including reasons for changing the original view.

At the deadline, I would score whether the specified event happened. A favorable interpretation cannot turn a failed forecast into a successful one.

Across many forecasts, I could check whether events assigned similar probabilities happen at similar rates. That is calibration. A few related predictions in one fashionable industry leave us with a sample too small and correlated to establish skill.

A simple way to evaluate a binary probability is the Brier score: square the difference between the assigned probability and the outcome, using one if the event occurs and zero if it does not.

For a 65% forecast, the score is 0.1225 if the event occurs and 0.4225 if it fails. Lower is better. A constant 50% forecast scores 0.25 for either outcome. We should compare a sequence of predictions with a reasonable baseline, report unresolved cases and avoid selecting only favorable results.

Calibration answers a different question: among many events assigned a probability near 65%, do roughly 65% happen? We need many sufficiently distinct events to assess that. One successful confident prediction can be luck.

The record should retain the original probability, with a date and reason for each update. Readers should be able to distinguish evidence that changed my view from a correction to the arithmetic. Changing a threshold after seeing the outcome creates a different forecast; it does not repair the original one.

A research routine we can actually use

I would start with a small number of consequential changes rather than a very large ticker list. For each change, the first task is to identify a customer and a spending decision likely to become observable.

Each week, I would read the relevant original technical releases, customer commitments and supplier results, then update a short evidence record. I want to detect changes in behavior. More mentions of the theme alone would not move my view.

A useful capability connected to funded adoption and a supplier with financially meaningful exposure would justify a deeper investigation. I would also examine a direct substitute or another business that could capture the same spending.

For each candidate, I would record why it could work, the strongest objection and what the valuation requires. Each update should identify which part of the argument the new evidence affects.

For AI medicine, I would start with the size and repeat rate of AI-related purchases, then measure external spending per useful candidate. I would also investigate ownership of experimental data, comparable pricing and returns on incremental manufacturing capacity. Answers to those questions are more likely to change my investment decision than another broad prediction that medicine will improve.

To finish the historical reconstruction, I need the financial information and share counts available in October 2025 for both attractive and disappointing candidates. That would let me test whether the early signals survived valuation scrutiny. So far, the chronology shows that the idea was discoverable. It does not show that the stocks were cheap.

The same questions apply to an adjacent theme. If a capability becomes cheaper, which customers can now afford to use it, and what else will they need? Which services become less valuable? I would then examine how quickly supply can respond, which listed businesses have material exposure and what their prices require. The answers will differ across industries.

I want the record to show my mistakes clearly enough that I can learn from them as evidence arrives.

The next large trade may come from a change we already recognize. Finding it requires following that change into customer spending, identifying a business that keeps the profit and buying at a price that leaves room for error.

What change are you watching today that could alter a customer's budget within the next twelve months?

Historical examples use information published by October 6, 2025. The prospective example uses a June 2026 baseline and an October 2027 evidence deadline.