Four corrections made the KISS signals more coherent even though their combined backtest return fell. A model should not keep a counting or weighting error because that error helped its historical performance.
The September 7, 2026 review used a January 2, 2020–September 4, 2026 reconstruction. The validation record separates construction fixes from hypotheses that still needed evidence.
The corrections adopted in KISS 2.4
Neutral sector signals now count as neutral, with symmetric 60% thresholds for bullish and bearish participation. An incomplete sector set is unavailable, rather than silently counted as neutral evidence.
Composite ETF baskets now use equal-weight return indexes over common observations. The old sum of share prices allowed a fund’s nominal share price to affect its influence. A stock split or arbitrary rescaling should not change a macro signal.
DBC trend uses adjusted returns, recognizing reinvested fund distributions. This measures an investable fund’s trend; it is not a spot-commodity inflation index.
Coverage is assessed across four equally weighted evidence domains. Losing an entire domain can no longer leave the model looking fully covered merely because other domains have many series.
| Construction, same sample | CAGR | Maximum drawdown |
|---|---|---|
| KISS 2.3 control | 15.11% | −8.07% |
| Sector correction alone | 15.32% | −8.07% |
| Equal-weight baskets alone | 14.54% | −8.69% |
| Adjusted DBC alone | 15.13% | −8.07% |
| Coverage correction alone | 15.16% | −8.19% |
| All four corrections | 14.69% | −8.82% |
The combined fixes were adopted for valid signal construction. They were not presented as a return improvement.
Ideas that did not clear the evidence bar
The review also tested calendar-consistent VAMS, simpler excess-return trends, hierarchical regime voting, real-yield/breakeven inputs, credit alternatives and a VIX stress overlay. These were distinct experiments, not ingredients combined after seeing which individually performed best.
VIX looked promising on the legacy construction: CAGR rose from 15.11% to 16.01%, with a smaller drawdown. The advantage weakened under delayed execution and conservative macro publication lags. On the corrected construction, the delayed-execution return advantage was only about 0.05 percentage points, and the publication-lag return fell below its control. VIX remained a shadow candidate.
The SOFR-minus-EFFR candidate worsened both return and drawdown in the initial comparison. The faster confirmation candidate also failed its tradeoff, as described in the confirmation study.
What these tests cannot establish
The publication delays were fixed stress allowances, not a reconstruction of every historical release vintage. The sample had already been inspected. Multiple candidates tested against the same history do not create independent out-of-sample evidence.
True constituent breadth also required historical memberships and delisted-stock coverage that the project did not have at the review date. Adding more ETF ratios would not solve that problem. The data-methodology study explains that boundary.
The useful distinction is between correcting how evidence is measured and adding a new trading rule. Only the latter needs to earn a performance claim.