Can a cap on stocks that move together improve returns?
We tested two ways to limit purchases of stocks that move together. Neither test gave us enough evidence to pick a cap.
English research index · Japanese original and update history
A portfolio can hold several stocks and still lose money across many positions at once if those stocks tend to move together. We used historical data to test a rule that limits purchases of such stocks. The study did not find a consistent improvement from introducing a cap. It also left open whether to use 25% or 30%.
The backtest was run on September 27, 2026. The figures are simulated returns. The conditions did not reproduce AlpAI's live trading environment closely enough to make an adoption decision.
What the cap limits
A 25% cap means that, after a purchase, existing positions plus the new purchase in the same group can make up no more than 25% of portfolio assets. The limit applies across the group. It cuts a purchase that would take the group over the cap and blocks additional purchases if the group is already at the cap. A group that moves above the cap as prices rise is not automatically sold under this rule.
Before each month we used the previous 120 trading days of prices to form 11 groups. The groups reflect how stocks moved together. We used them as an exploratory substitute because historical industry classifications were incomplete.
Two questions, two comparisons
Both tests covered 2020–2025 and used ten starting dates spaced five trading days apart.
| Comparison | Universe and conditions | Question |
|---|---|---|
| No cap, 25%, and 30% | 513 stocks. Three stocks with short histories were excluded throughout. | Did adding a cap improve returns? |
| 25% and 30% | 516 stocks. Eligibility for new purchases was set each month based on available history. | How did the two caps compare under the same monthly eligibility rules? |
We could not produce a no-cap comparison under the same monthly eligibility conditions in the 516-stock test. The no-cap result below comes from the 513-stock test. Compare results within each test.
The 513-stock test against no cap
The annualized return differences are averages of comparisons made at each of the ten starting dates.
| Cap | Average annualized return difference versus no cap | Starting dates ahead of no cap |
|---|---|---|
| 25% | +0.43 percentage points | 5 of 10 |
| 30% | −1.11 percentage points | 2 of 10 |
The 25% cap had a positive average difference and finished ahead on half the starting dates. The 30% cap had a negative average difference. Taken together, these results offer no consistent case for adding a cap under the conditions tested.
The starting dates overlap substantially. Five of ten is a count of overlapping simulations. It does not estimate a 50% chance of beating the no-cap rule in the future.
The 516-stock test comparing the two caps
| Comparison | Average annualized return difference | Starting dates with 25% ahead |
|---|---|---|
| 25% minus 30% | +1.92 percentage points | 9 of 10 |
The 25% cap came out ahead in this test. This comparison had no no-cap baseline, and its ten starting dates do not provide ten independent observations. The result alone is insufficient to choose 25% or to adopt a cap.
We also compared maximum drawdown, the largest fall from a prior peak. The two caps had nearly the same average maximum drawdown. For some starting dates, the 25% cap made the drawdown worse, by as much as 1.70 percentage points.
What we would test next
A stronger test would run all three choices on the same eligible stocks under the same trading conditions. For now, neither comparison gives us a clear reason to adopt this rule or settle on either cap.
The study tested groups based on historical price behavior. It does not settle the value of industry caps or diversification more broadly. Some months still need point-in-time data availability checks. The simulation also differs from live trading in stock eligibility, purchase conditions, measured costs, and the timing of decisions. Parts use same-day prices, which leaves open whether those trades could have been made at those prices.
Updates to the Japanese original
- September 29, 2026. Clarified that the cap includes existing positions and separated the 513-stock and 516-stock comparisons. Added context about drawdowns across starting dates. The figures and conclusion were unchanged.
- October 3, 2026. Clarified the overlap between starting dates and noted that, under this setup, a difference of roughly 2.5 annualized percentage points or less was difficult to interpret. The +1.92-point result did not establish that 25% was better.
- October 4, 2026. Revised the framing so the 25% result would not read as a preference for that cap. The period, conditions, and results were unchanged.
- October 8, 2026. Reordered the explanation and stopped using approximately 2.5 points as a mechanical cutoff for superiority. The period, conditions, and results were unchanged.
For general background on reading investment performance, see the SEC's investor guide. It does not validate this backtest.
The figures describe a historical simulation. They do not guarantee future results or recommend either cap.