Applied AI / Worked case /
The test behind
the number.
A calculation can be reproducible while the comparison still needs arguing. Here is one example, with the cases and the assumptions visible.
What changed when we checked
Seven matches.
Then none.
Following the 22 October 2026 semiconductor sell-off, an independently rerun SOXX study found seven completed historical episodes under a two-condition rule. Their median price return 60 trading sessions later was +26.2%.
We added one condition: the signal close also had to be above its 200-day average, as it was on 22 October. None of the completed historical cases met that condition.
The extra trend condition is exploratory. Zero matches does not invalidate the broader sample or predict what happens next. It exposes an assumption behind transferring a historical result to a different setting.
The same starting point, different definitions
What counted as a match?
Each row selects its own events and applies a 60-session exclusion window. The added trend condition is applied before event selection.
| Selection rule | Completed events | Median 60-session return |
|---|---|---|
| RSI only | 24 | +3.7% |
| Distance from 20-day average only | 10 | +11.9% |
| Both conditions | 7 | +26.2% |
| Both + above the 200-day average | 0 | No estimate |
RSI(14) threshold: 30.802174 or lower. Distance from the 20-session simple average: −12.905205% or lower. These thresholds come from the 22 October 2026 observation, not from a rule fixed before the historical sample.
The full main sample
The finish and the path.
A positive eventual return can hide a difficult journey. The median lowest subsequent close was 16.5% below the signal close.
| Signal close | 60-session return | Lowest subsequent close vs signal | Signal vs 200-day average |
|---|---|---|---|
| 2001-09-17 | +27.4% | -19.3% | Unavailable |
| 2002-06-11 | -32.8% | -32.8% | -20.6% |
| 2002-09-20 | +26.2% | -16.5% | -46.8% |
| 2008-10-07 | -12.2% | -30.9% | -33.5% |
| 2011-08-08 | +12.7% | -1.7% | -20.7% |
| 2020-03-12 | +43.6% | -6.6% | -15.1% |
| 2025-04-03 | +39.5% | -9.1% | -23.6% |
| Median | +26.2% | −16.5% | 6 below; 1 unavailable |
The two medians describe different statistics, not a forecast of one representative path. The loss measure uses subsequent daily closes over 60 sessions; it is not an intraday or peak-to-trough drawdown.
Download all seven cases (CSV)A practical connection
Test the work.
Then test the test.
In his 2025 Jane Street engineering talk, John Crepezzi describes a Code Evaluation Service. A model proposes a change. The service applies it to a known-good code revision, runs compilation and tests, and returns the outcomes to training. Held-out examples support evaluation.
The useful connection is external verification of proposed work. The boundary matters too: a compiler cannot determine whether a market comparison is economically appropriate. That judgment needs its own evidence.
This case makes no claim that AI beat a bank’s risk system, that Jane Street replaced its financial-risk engines, or that the historical return predicts the next one.
Watch the relevant Jane Street explanationTry it in your own work
Ask for the
assumption.
Reproduce the result. Change one economically relevant assumption. Show every case included and excluded, and explain what you can no longer conclude.
Use the same decision and acceptance conditions when comparing workflows. A passing numerical check establishes the arithmetic; source quality, interpretation and predictive usefulness remain separate tests.
Changing more conditions can overfit the past. A genuinely predictive claim needs a rule fixed in advance and assessed on observations it has not already seen.
Data, method and limits
The study used 6,357 daily SOXX observations from 13 July 2001 to 22 October 2026. Yahoo Finance close prices are split-adjusted, not dividend-adjusted total returns. RSI uses a 14-change arithmetic seed followed by Wilder smoothing; simple moving averages use the trailing 20 and 200 closes including the signal date.
Select the first qualifying close, exclude the next 60 sessions, and require a complete 60-session future window. Use exact thresholds; round only for display. The current episode is not part of the forward-return sample. Dividends, costs and executable-entry assumptions are excluded; ETF composition changed over the period.
This is a dated assumption audit, not a validated trading strategy or a head-to-head benchmark of AI against institutional risk systems.