Actionable insights — Buy-Hold-Sell List: May 2026

1. Rate quality and price on two independent axes, and let them disagree

1. Give every holding a conviction rating about the business — how sure you are it will still be excellent in ten years.
2. Give it a separate rating about the price — what return the current valuation implies.
3. Never let one contaminate the other. A wonderful business at a silly price is a Hold, not a Buy; a mediocre business at a very cheap price can be a Buy without being a conviction.
4. Publish both, so the disagreements are visible and reviewable.
5. Use the combination for sizing: high conviction plus a large gap gets the money.

• The two axes quietly merging over time. If every high-conviction name is also a Buy, one of the columns has stopped doing work.
• The absence of a rung: nothing in the portfolio is rated SELL, on a list called Buy-Hold-Sell.

2. Track your own buy count over time and read it as a market signal

1. Fix the universe and the screening criteria so the count is comparable month to month.
2. Record how many names clear your buy threshold each month.
3. Read the series, not the level: a record high count means your style has de-rated, and a record low means it has become expensive.
4. Corroborate with an external factor measure before drawing a conclusion, so you are not just reading your own model's drift.

• The count rising because the model's assumptions drifted rather than because prices fell — the exit multiples are inputs you choose.
• "Opportunity of a century" language. A record buy count is a factor observation; the timing of the reversal is not in it.

3. Require agreement across three independent valuation methods before calling something cheap

1. Run each candidate through a multiple-based test (forward PE against its own history), a return-decomposition model (growth + yield ± re-rating) and a reverse DCF (what growth the price implies).
2. Publish all three answers, including where they disagree.
3. Treat the intersection — undervalued on all three — as the real list.
4. Where they conflict, work out which assumption is doing it, usually the assumed exit multiple.

• The exit multiple carrying the result. GOOGL reaches a Buy at $382 on a 25x assumed exit, sixteen days after the stated entry was 18x, i.e. $210. Same company, different assumption.
• Methods that are not actually independent — all three here use the same forward earnings estimate.

4. Keep an explicit veto over your own screen, and disclose when you use it

1. Let the model produce the rating mechanically, without adjustment.
2. Apply judgement separately, as a documented override, rather than by tweaking the inputs until the answer changes.
3. State the reason for the override in one line, so it can be revisited.
4. Recognise that a published list a reader might follow mechanically now differs from what you would do.

• The gap between what the list says and what the author does. That gap is where a subscriber gets hurt.
• "Too hard" being applied selectively — it excuses Adobe but not the 49 other names whose exit multiples embed the same technology assumptions.

5. When you cannot pick the winner, buy the input every contestant needs

1. Admit the winner is unforecastable, and say so rather than hedging with a basket of contestants.
2. Map the value chain and find the layers every participant must buy from — compute, power, land, cooling, connectivity.
3. Prefer layers with physical scarcity and long permitting or construction lead times, which cannot be competed away quickly.
4. Check that the supplier is actually capturing the economics, with disclosed contracts and capacity, not just adjacency.
5. Verify with signed commitments rather than management ambition.

• Shovel sellers that become cyclical when the rush ends. Data-centre capacity is a build cycle, and it can be overbuilt.
• Adjacency without economics — plenty of companies are "AI-exposed" without pricing power.

6. Measure the concentration of your benchmark before calling it diversified

1. Find the top-ten weight and the largest sector weight in the index you are compared to.
2. Ask whether the largest names are exposed to the same driver, which is a stronger form of concentration than sector labels capture.
3. Compare to the long-run average concentration for context.
4. Then judge whether your own portfolio is genuinely diversifying away from it, or duplicating it.

• Concentration measures that flatter — the top ten at 38.5% here versus the "nearly 40%" quoted three days later.
• Owning the same driver through a different door: several names on this Buy list are software businesses facing the same AI question as Adobe.

7. Reconcile your own published tables before you publish them

1. Cross-check the counts in the prose against the rows in the table.
2. Cross-check a name appearing in two tables for the same rating.
3. Cross-check the adjustments you insisted on elsewhere — if you argued a multiple must be restated after stock-based compensation, use the restated figure in the model.
4. Where two numbers must differ (different date, different basis), say why in a footnote.

• Adjustments that appear in the prose and vanish in the model. The model is what drives the rating.
• A stale label surviving a template — both performance graphics in the 3 May issue are titled "March 2026".

8. Upgrade a rating with the open objections still stated, not resolved

1. When the price moves enough to change a rating, change it — and keep the unresolved business questions in the write-up.
2. Say which of the two changed: the price, or your view of the company.
3. Size accordingly. A valuation upgrade on an unresolved business question is a smaller position than a conviction upgrade.
4. Set the evidence that would settle it, and the date you will look.

• The short-form version of the same idea dropping the caveats. Cadence matters: FICO appeared three times in five days, twice without the risk.
• A rating change presented as new information when only the price has moved.