EXPERIMENT ARCHIVE
Failure is signal.The gate stays closed.
Salarium preserves attractive hypotheses that failed governance alongside the decisions that shaped the locked release. Nothing is promoted because one aggregate number looks good.
LEADING VARIANT20% oilresearch leader, not selected policy
ANNUALIZED NET48.1%50.1% cash-yield comparator
MAX DRAWDOWN-33.4%-49.2% comparator
NET SHARPE1.4971.335 comparator
FROZEN ACCEPTANCE GATES8 / 9 passed
Maximum drawdownPASS
Expected shortfallPASS
Recovery timeFAIL
Return dragPASS
Net SharpePASS
Yearly consistencyPASS
Holdout drawdownPASS
Holdout SharpePASS
25 bp cost stressPASS
DECISION / NON PROMOVETUR
Promising sample.
Insufficient hedge.
The leading 20% oil comparator improved maximum drawdown by 15.8% and Sharpe by 0.162, but reduced the longest recovery by only 3.6% against a frozen 20% requirement.
Its strength is concentrated in the inflationary sample. It is not reliable enough to alter the release architecture.
PRE-SPECIFIED STRESS WINDOWS / ETF TOTAL RETURN2008—2022
WindowSPYGLDUSOTLT
Global financial crisis-46.4%10.8%-69.2%13.0%
US debt downgrade / euro stress-17.5%11.1%-19.9%33.4%
Q4 2018 equity selloff-19.2%6.6%-41.8%5.2%
COVID-19 liquidity shock-33.7%-3.6%-56.4%14.2%
2022 inflation/rate bear market-24.5%-7.3%29.4%-29.3%
ETF proxies, not contract-level futures. Integrated Salarium evidence covers 139 simulated out-of-sample rebalances from 2021–2026; 2026 is partial. The fair comparator adds Treasury-bill yield to unused capital. No live performance.
LOCKED RELEASE LEDGER / I—VI
EXP 01ACCEPTED
Keep the Liquid-500 portfolio universe
- Hypothesis
- Did expanding the portfolio universe to roughly 2,000 names improve out-of-sample performance?
- Result
- The broader point-in-time universe reduced drawdown in the risk-scaled policy, but materially weakened annualized return, Sharpe, and ranking IC. More names did not create a better portfolio model.
- Decision
- Retain Liquid-500 for Salarium 1.0 portfolio construction; keep broad coverage as a separate discovery and research funnel.
- Evaluation period
- 2021–2026
- Source artifact
- reports/experiments/broad_vs_liquid500_walkforward.csv
- Commit / updated
- 7c730e84c409 · Aug 11, 2026, 11:14 PM UTC
EXP 02ACCEPTED
Predict 20 days; rebalance every 10
- Hypothesis
- Was the original five-day target and five-day rebalance cadence too short and too active?
- Result
- Separating prediction horizon from trading cadence showed that Salarium's signal is slower-moving. The 20D model traded every 10 days improved both return and Sharpe versus the original 5D/5D design.
- Decision
- Lock a 20-trading-day model horizon and 10-trading-day rebalance cadence.
- Evaluation period
- 2021–2026
- Source artifact
- reports/experiments/horizon_rebalance_static_matrix.csv
- Commit / updated
- 7c730e84c409 · Aug 11, 2026, 11:14 PM UTC
EXP 03ACCEPTED
Concentrate on the Top-10
- Hypothesis
- Could a broader 20–75 name portfolio preserve alpha while reducing risk?
- Result
- Additional breadth reduced volatility and turnover, but diluted return faster than it improved risk-adjusted performance. The model's useful alpha remained concentrated near the top of the ranking.
- Decision
- Keep Top-10 concentration with a rank-15 persistence buffer; manage joint risk through covariance rather than indiscriminate breadth.
- Evaluation period
- 2021–2026
- Source artifact
- reports/experiments/portfolio_breadth_static.csv
- Commit / updated
- 7c730e84c409 · Aug 11, 2026, 11:14 PM UTC
EXP 04ACCEPTED
Replace standalone risk with joint risk
- Hypothesis
- Could covariance-aware construction preserve concentrated alpha while reducing redundant correlated risk?
- Result
- A 60D Ledoit-Wolf maximum-diversification portfolio improved Sharpe and Sortino while modestly improving drawdown versus inverse-volatility weighting. The optimizer completed without fallback in the selected configuration.
- Decision
- Use 60D shrinkage maximum diversification as the primary risk anchor; retain minimum variance as the defensive comparator.
- Evaluation period
- 2021–2026
- Source artifact
- reports/experiments/covariance_portfolio_overall.csv
- Commit / updated
- 7c730e84c409 · Aug 11, 2026, 11:14 PM UTC
EXP 05ACCEPTED
Give the signal a governed 25% vote
- Hypothesis
- Should conviction influence weights after Top-10 selection and covariance optimization?
- Result
- A 25% signal blend increased the balanced mandate's simulated return while leaving overall Sharpe nearly unchanged. Higher blends continued to raise return but progressively increased volatility and drawdown.
- Decision
- Blend 25% signal-aware weights with 75% covariance-risk weights under the 18% single-name cap.
- Evaluation period
- 2021–2026
- Source artifact
- reports/experiments/signal_aware_covariance_overall.csv
- Commit / updated
- 7c730e84c409 · Aug 11, 2026, 11:14 PM UTC
EXP 06ACCEPTED
Cap leverage; never force it
- Hypothesis
- Did the evidence justify using portfolio exposure above 1.00x?
- Result
- The selected portfolio and exposure policies did not require leverage above 1.00x in the committed evaluation. The risk layer found more value in de-risking than in borrowing additional capital.
- Decision
- Retain a hard 1.25x governance ceiling as permission—not a target—and keep the selected mandate unlevered unless future risk evidence earns additional exposure.
- Evaluation period
- 2021–2026
- Source artifact
- reports/experiments/signal_aware_covariance_overall.csv
- Commit / updated
- 7c730e84c409 · Aug 11, 2026, 11:14 PM UTC