Does giving AI more time to think improve stock selection?
Higher AI reasoning effort lifted a simulated account’s ending balance by $244.81. With the number of picks held constant, the higher-effort picks did worse against the benchmark.
English research archive · Japanese original
We changed the reasoning effort on the same 100 earnings reports. The higher-effort simulated account finished $244.81 ahead of the baseline and cleared the $100 threshold we had set for its ending balance.
Did the model pick better stocks? In follow-up checks, the added picks lagged SPY, an ETF that tracks the S&P 500, on average over ten trading sessions. When we compared 25 picks at each setting, the higher-effort picks also earned a lower average return relative to SPY. Those checks help explain what the account result can tell us about stock selection.
The October 27, 2026 study used earnings from a single quarter. Seventeen decisions changed, so the sample gives limited evidence about how either setting would perform in another quarter.
The experiment
- We selected 100 of 439 earnings reports for the second quarter of 2026 using a rule fixed in advance.
- We ran each report once at medium reasoning effort (the baseline) and once at higher reasoning effort. The model, source material and instructions were the same.
- The simulated account started with $10,000, could hold up to eight positions, held positions for ten trading sessions and deducted trading costs. Only the 100 selected reports switched to higher effort. The other 339 kept their baseline readings.
- We recorded the design and pass criterion before running the test. The higher-effort account had to finish at least $100, or 1% of starting capital, ahead of the baseline.
The simulated account finished ahead
The baseline account finished at $13,001.90. The higher-effort account reached $13,246.71, a difference of $244.81.
Follow-up: which stocks did the model add or drop?
The table shows after-cost stock returns minus SPY returns over ten sessions from the study’s specified starting point. The figures are in percentage points.
| Change at higher effort | Reports | Average difference versus SPY | Median difference |
|---|---|---|---|
| Switched to buy | 13 | −1.15 | −3.90 |
| Switched away from buy | 4 | +1.76 | −0.73 |
| Buy at both settings | 21 | +6.00 | — |
The four dropped names had a positive mean and a negative median. With so few observations, those summaries can tell different stories about the same group.
What happened with 25 picks at each setting?
The baseline marked 25 reports as buys. We compared them with the 25 highest-scoring reports at higher effort.
| Setting | Average ten-session difference versus SPY for the top 25, pp | Rank correlation of score and ten-session difference across all 100 |
|---|---|---|
| Baseline | +5.32 | 0.181 |
| Higher effort | +3.01 | 0.142 |
A simulated account using only the top 25 at higher effort finished at $12,948.55, about $53 below the baseline. In a separate check, keeping the original buy decisions and substituting only the scores from the higher-effort run improved the account by about $167. The scores affected which stocks got an available slot and when, including the ordering of candidates that appeared on the same day. This helps explain how the account could finish ahead while the stock selection checks weakened.
Where the decisions changed
The calculations for sales, profits and valuation multiples were nearly the same in the 17 cases that changed. The average score across all 100 moved from 49.0 to 48.8. The valuation labels changed more. Across all 100 cases, the model assigned an “uncertain” label to 22 at baseline and 11 at higher effort. The number labeled “cheap” rose from 30 to 38. In eight of the 13 cases that switched to buy, the label changed from “uncertain” to “cheap.”
| Change at higher effort | Median return over 20 sessions before the release | Median reaction at the open after the release |
|---|---|---|
| Switched to buy, 13 reports | −6.7% | +3.1% |
| Switched away from buy, 4 reports | +14.4% | +6.1% |
At higher effort, the model tended to interpret earlier declines as a sign that a recovery was still missing from the price. It more often read earlier gains as evidence that growth was already priced in. In this quarter, that tendency did not translate into stronger ten-session returns for the newly selected stocks. The test leaves the cause open.
Context: archived baseline readings
A separate review summarized 2,331 archived baseline readings from the same quarter. This different sample describes the baseline setting. Its group averages cannot measure the effect of switching settings in the 100-report experiment.
| Valuation label | All reports | Reports with a ten-session result | Average difference versus SPY, pp | Median difference, pp |
|---|---|---|---|---|
| Cheap | 210 | 207 | +4.14 | +2.85 |
| Fair | 1,399 | 1,391 | −1.90 | −2.18 |
| Expensive | 152 | 152 | −1.25 | −2.51 |
| Uncertain | 570 | 567 | +1.66 | +0.52 |
Fourteen reports lacked a ten-session result and were excluded from the averages. Of 179 buys, the model labeled 173 “cheap.” The 170 with results averaged +4.75 points versus SPY, with a median of +3.05. The other buy categories each contained three reports or fewer, which limits comparisons among valuation labels.
In the 100-report experiment, higher effort raised the buy count from 25 to 34 even though the score threshold stayed the same. The tests leave the effects of the earnings news, market reaction and the model’s valuation labels intertwined.
What we would check before using it
The higher setting used roughly six times as much reasoning and cost more. We kept the baseline setting in live use at the time of the study. The research plan called for recording higher-effort readings on third-quarter reports without trading on them, then checking in late December whether the simulated account improved by at least 1% of starting capital and whether changed decisions showed a positive ten-session difference versus SPY.
Researchers at Anthropic have documented examples where extra reasoning worsened answers. This experiment adds an observation from earnings analysis. Its one-quarter design leaves open whether the same mechanism was involved. Inverse Scaling in Test-Time Compute
If an AI change improves a portfolio simulation, check its picks at the same trade count, the outcomes of the names it adds and drops, and whether the order of trades changes the account result. In this study, only the final account balance was a pre-set pass criterion. The stock-level comparisons were follow-up diagnostics.
Updates to the Japanese original
- November 2, 2026. Revised the structure and wording. The periods and numerical results stayed the same.
- November 2, 2026. Clarified that ten-session results were available for 2,317 of 2,331 archived readings and 170 of the 173 buys rated cheap. Specified percentage points as the unit of the differences.
AlpAI Lab publishes research and records from its own investing. We do not give personalized investment advice.