Close-to-close, exact drift
Daily total returns update every position. Portfolio weights drift naturally between trades. Scheduled rebalances occur at quarter-end unless a specific research variant states otherwise.
06 · Credibility layer
Rules before
results.
2 engines
The frozen tournament and an independent replication must agree. A compelling chart is not accepted as evidence until the mechanics reconcile.
Backtest contract
What the engine does.
What it does not.
The goal is not to make the backtest look clean. It is to make every assumption visible enough that an investor can disagree with it intelligently.
Observed first, traded later
Signals end at t−1. The decision is implemented at close t and first affects return on t+1. No strategy is allowed to react to a close it has not yet observed.
Every dollar bought and sold
Cost at a trade is 0.001 or 0.002 multiplied by the sum of absolute weight changes. A full rotation of $1 million creates $2 million of chargeable notional.
SGOV, never borrowed capital
Released risk goes to SGOV. All final weights are nonnegative, exposure never exceeds 100%, and the research shortlist contains no negative-cash or portable-alpha implementation.
Stale execution is deferred
If an ETF being traded has no fresh close, the whole requested trade waits. This affected BTAL on 29 March 2018 and SHYG on 30 June 2020, 30 June 2022 and 30 June 2026 in paths where each ETF was held.
Formation and liquidation separated
Deployment from external cash is 100% buys; final liquidation is 100% sells. Each costs 10 or 20 bp and is shown separately from ongoing operating comparisons.
Late-launch ETFs
Never silently
treated as zero.
This point is explicit because it materially changes the interpretation of the ten-year result.
Proxy-spliced, 2016–2026
01If a selected ETF launched after the backtest began, its intended exposure is replicated with the closest available ETF or the history of its tracked index where possible.
02The portfolio is not assigned a fictional 0% return or left uninvested simply because the live wrapper did not yet exist.
03Because proxy-to-live instrument migrations are not yet mapped on every date, full-history execution costs are labeled research-only.
Clean restart, May 2022–2026
01Every portfolio begins from its exact policy target on the common-live start. Prior holdings, drift and overlay state are discarded.
02Pre-start observations exist only to warm up moving averages and volatility signals; they do not contribute portfolio returns.
03The full and fresh-native rankings are shown together so a result dependent on synthetic early history cannot hide.
Frozen research design
Small grid.
No parameter hunt.
Candidate holdings and signal thresholds were declared before the run. Robustness periods test stability; they are not used to pick whichever parameter happened to win.
Holdings
Current basket, QQQ only, QQQ + GRID, QQQ + GRID + BUG, and a separate QQQ + SOXX sensitivity.
→
Governor
Static, Soft, Hard, three-state, and three-state plus weekly or monthly nonlevered volatility brake.
→
Acceptance
Compare return, volatility, drawdown, rolling three-year outcomes, regimes, turnover and both cost tariffs.
Outcome: the current specialist basket remained the return control; QQQ-only is the simplicity challenger; the low-frequency Soft governor is the only overlay retained for paper testing. The complex three-state and both volatility-brake variants were rejected.
Independent audit
Checks before conclusions.
Passing checks do not prove future performance. They establish that the reported results follow the stated rules and arithmetic.
01PASS
Lagged signals
Signals use information through the prior close. A close-t trade first changes the next session’s return.
02PASS
No leverage
All portfolio weights remain nonnegative and sum to 100%. The brake multiplier is capped at 1.0.
03PASS
Two-leg costs
Sell + buy equals gross traded notional on every event; 10/20 bp formulas reconcile at machine precision.
04PASS
Static replication
A second independent simulator reproduces every frozen static path within approximately 1e−16 daily return.
05PASS
Stop replication
All 12 existing Strategy Lab stop paths reconcile before cost repricing within approximately 2e−16.
06PASS
Stale prices
Affected quarter-end trades are deferred until all ETFs being traded have a fresh executable close.
07PASS
Fresh-native restart
Holdings and rule state restart from target; proxy-era drift is not inherited into the native comparison.
08PASS
Deterministic payload
Two consecutive website-data rebuilds produced the same checksum and passed all critical validations.
Limitations that remain
Precision without
false certainty.
The website is intentionally explicit about what has not been solved.
Proxy execution
Index history can inform return research but is not always an investable instrument. Exact migration dates and costs still require a complete instrument-by-date map.
Trading frictions
The 10/20 bp tariff is modeled exactly as supplied. Taxes, bid/ask spread beyond the tariff, market impact, custody and cash-flow timing are excluded.
Selection risk
Even a frozen ten-year tournament is in-sample research. The retained Soft rule must be paper-traded and monitored before any live claim of edge.