A promising change.
A higher bar for evidence.
Recent work has focused on testing a candidate signal change and making the research record easier to review. Some historical results improved. The evidence did not establish a consistent advantage.
This update reviews selected work documented in September and October, with source material through October 17. It reports research progress; the earlier market notes remain dated observations in the archive.
01 / What we learned
An exploratory comparison tested an existing signal approach against a candidate extension, using the same inputs within each comparison. In one historical window, the two versions produced the same portfolio outcomes. In a separate full-year historical test, the candidate showed a modest improvement, concentrated in a small part of the stock universe.
Controlled artificial paths gave mixed results: some outcomes improved, some deteriorated, and others were unchanged. This helped reveal where the change affected decisions, without establishing how often those conditions occur in real markets.
The October 17 review retained the candidate as research and did not adopt it into the main model. A local improvement is a reason to investigate further, rather than evidence of a dependable advantage. No statistical significance claim is made.
02 / How we tested
Compare versions on shared inputs
Within each historical comparison, both versions used the same data and accounting assumptions. Different historical windows were reviewed separately; their absolute results were not treated as a controlled comparison.
Challenge the result
The full-year study also examined higher assumed costs and delayed execution. Artificial scenarios explored contrasting trends, reversals and choppy paths. Neither sensitivity checks nor artificial paths can substitute for unseen observations.
Preserve the evidence
The October study was archived with its comparison inputs, assumptions, results and limitations. Separate work documented how saved simulation records can be audited while keeping missing observations and operational failures visible.
03 / What remains unresolved?
The main questions are whether the observed improvement extends beyond the historical sample, whether it is spread across enough instruments and market conditions, and whether the parts of the hypothesis with little or no observed activity can be evaluated.
Those questions remain open. A reproducible backtest or a working simulation pipeline establishes a narrower fact than investment effectiveness. Any stronger conclusion would require separately defined validation and an adequate independent sample.
Keep the claim within the evidence.
Editorial date: October 21, 2026. Source cutoff: October 17, 2026. This is a research-process update, with no current market-price cutoff or new trading signal.
Source provenance: the internal signal-extension feasibility archive dated October 17, its full-year historical comparison, and research-validation documentation dated September 29 and October 13. These are internal research records, summarized here rather than published as underlying datasets or an independent external audit.
Limits: selected historical windows, a small universe chosen with hindsight, retrospectively adjusted prices, simplified costs and fills, and uneven coverage of the hypothesis. Artificial scenarios are constructed examples. The work summarized here does not establish out-of-sample performance or a live investment track record.
Proprietary parameters, model construction, execution rules, account information and internal reports remain private. No performance figures are published in this update.