Research note
What a backtest can—and cannot—establish
A benchmark comparison is the beginning of the review. Selection, dispersion and a failed sector test add the context.
Research snapshot Oct 15, 2026. These results are simulations.
A result is evidence for a specific test
The question is whether a learned trading policy performs better than a matched benchmark during a specified evaluation period. A return number only becomes interpretable alongside the dates, universe, costs, comparison and selection process.
For the latest tested stock group in each sector below, we show the median across valid full-window tests, the complete return range and the count ahead of SPY. The median describes a set of simulations; it is not a combined account return.
Read the spread of results
Technology
78 tests+90.9%median backtest
S&P 500 proxy · SPY +12.0%
Jan 2 – Aug 31, 2026
Retrospective stock selection
- Difference vs SPY
- +78.9 pp
- Tests ahead of SPY
- 78 / 78
- Full return range
- +29.3% to +158.9%
- Median maximum drawdown
- −18.1%
- Stock group selected
- 2026-10-13
Energy
12 tests+29.3%median backtest
S&P 500 proxy · SPY +12.0%
Jan 2 – Aug 31, 2026
Retrospective stock selection
- Difference vs SPY
- +17.3 pp
- Tests ahead of SPY
- 10 / 12
- Full return range
- +1.8% to +49.5%
- Median maximum drawdown
- −7.2%
- Stock group selected
- 2026-10-15
Healthcare
10 tests−12.6%median backtest
S&P 500 proxy · SPY +12.0%
Jan 2 – Aug 31, 2026
Retrospective stock selection
- Difference vs SPY
- −24.6 pp
- Tests ahead of SPY
- 0 / 10
- Full return range
- −19.2% to −6.7%
- Median maximum drawdown
- −21.8%
- Stock group selected
- 2026-10-15
Each comparison uses the same evaluation timestamps and SPY return within that sector cohort. Sector stock groups have different selection dates. A completed experiment can be a losing test; completed work and successful investment performance are different measurements.
Selection changes the conclusion
The stock groups shown were selected after the displayed historical evaluation period. Multiple model configurations have also been evaluated on the same holdout. Those choices can inflate apparent performance. The figures support further investigation, not a claim of independently verified out-of-sample or live results.
A useful next test freezes the selection rule before seeing new outcomes, includes trading assumptions, and compares against simple alternatives. Replication across seeds and fresh periods helps identify results that depend on a fortunate run.
An original study with a different answer
The published sentiment study covers nine stocks and 346 trading days. In its within-stock, 60-day-relative news-sentiment comparison, the same-day t-statistic is +6.9; the next-day value is −0.4. That distinction matters: a variable can describe a price move without providing a next-day lead.
Read the full study, charts and method →
A finding that removes an attractive idea can be useful research. The next decision becomes more focused when we preserve those results instead of publishing only the favorable charts.
What is published here
Sector summaries require saved successful evaluation metrics, completion history and no later integrity invalidation. Figures remain exploratory. Live portfolio figures are separate on the portfolio page.
Research summary and timestamps → · Macro and policy snapshot →