Preserve capital. Then grow.

The architecture

1. Gather and reconcile point-in-time company filings, earnings calls, institutional activity, analyst revisions, brokerage data, market structure, rates, volatility, and macroeconomic evidence in an owned research lake.
2. Construct proprietary company, sector, and regime features with production data pipelines that preserve timing, provenance, and revision history.
3. Run governed AI inside Corbex's fundamental-analysis harness using versioned prompts, fixed schemas, source citations, validation gates, and stability checks.
4. Combine governed features through versioned machine-learning models, then apply explicit rules for ranking, concentration, sizing, and defensive exposure.
5. Implement portfolio changes in highly liquid US-listed securities through opening-auction orders, then retain the complete research-to-production lineage for review.

Strategy characteristics

Core exposure

Concentrated US-listed growth equities

Typical long book

Approximately 10 to 20 companies

Defense

Broad-market or sector ETF shorts

Decision cycle

Daily research and manager-approved market-on-open execution

Methodology and assumptions

Decision cycle: Daily research and manager-approved market-on-open execution

Execution: Market-on-open portfolio rebalance timing.

Trading costs: 10 basis points per unit of turnover.

Financing costs: 75 basis points annualized for modeled ETF borrow exposure.

Benchmark: QQQ - Buy-and-hold total return with dividends reinvested (dividend-adjusted closes), rebased to $1.00 on 2018-01-09 and aligned to every Arbiter session through 2026-08-04. It provides a same-window US large-cap equity comparison and is not the index itself.

Data window: 2018-01-09 through 2026-08-04, covering 2,153 trading sessions.

Simulated returns reflect Market-on-open portfolio rebalance timing, 10 basis points per unit of turnover, and 75 basis points annualized for modeled ETF borrow exposure. No fund fee is applied in this simulation.

The research evidence

Historical simulation - frozen and reproducible

Validation evidence

Selection-adjusted method

Bailey–López de Prado Deflated Sharpe Ratio - The Deflated Sharpe Ratio adjusts the selected complete-strategy Sharpe for the measured search family, sample length, cross-path dispersion, skewness, and kurtosis. It validates the complete strategy within the search family and does not independently validate the defensive overlay.

Strategy complexity

Complexity score: 5

Controlling for selection bias

The discovery program evaluated 183 candidate formulas across 169 distinct economic paths before this strategy was registered. Reporting the best result of a large search as if it were a single experiment is how backtests flatter themselves, so the record is judged with the Deflated Sharpe Ratio: the observed Sharpe ratio must clear a hurdle raised by the breadth of the search, the length of the record, and the non-normality of returns. Sharpe ratios use a 0% risk-free rate. Annualized return compounds at 252 trading sessions per year. Sortino downside deviation uses a 0% target averaged over all sessions.

The record clears that hurdle under every disclosed treatment - including the registered punitive treatment, which charges the entire search against the strategy.

Validation windows

Cost sensitivity

Overlay diagnostic

Does not pass 95% incremental DSR - The complete simplified strategy passes the stage-specific DSR within its immediate search family. The defensive overlay does not independently clear the 95% threshold; its case remains bounded risk logic, drawdown behavior, temporal diagnostics, and prospective observation rather than a standalone-alpha claim.

Limits

• The 2023–2026 extension was already inspected during final strategy selection and is not a pristine out-of-sample holdout.
• The complete-strategy DSR does not independently validate the defensive overlay or establish standalone defensive alpha.
• The calculation covers the immediate 183-formula defense family and 169 distinct economic paths, not the full research lineage.
• The frozen record does not include the 2 bps daily selection path; its scenarios conservatively reuse the registered discovery skewness and kurtosis with the reported 0.03491 selection-frame Sharpe dispersion.
• Cost sensitivity is reported from the two published frames - the gross frame and the registered 10 bps trading / 75 bps borrow net frame - over the full record and the discovery window; intermediate cost levels are not separately simulated.
• Results are gross of prospective investor fees, vehicle expenses, taxes, financing outside the declared borrow assumption, and other implementation differences.

What validation can and cannot prove

These tests establish that the simulated record is unlikely to be an artifact of search: the mechanics are frozen under cryptographic locks, the data is point-in-time, discovery and extension windows are reported separately, and the result survives punitive cost and selection treatments. Historical simulation cannot establish execution against real liquidity under real operations. We research, execute and monitor Arbiter daily. This research evidence does not establish live validation or promise future results.

Why this backtest is reproducible

The daily series is frozen, and every displayed statistic reproduces directly from it under a hash-checked release manifest. Results reflect the registered cost model - 10 basis points per unit of turnover and 75 basis points annual borrow. No fund fee is applied in this simulation. The Deflated Sharpe analysis charges the result for the measured search that produced it. Reproducibility of a historical simulation does not establish its implementation assumptions in live operation. We research, execute and monitor Arbiter daily.

Risk factors

• Concentration can amplify company, factor, and sector losses.
• A defensive sleeve can lag in sharp reversals or fail to offset long-book losses.
• Research selection and reused historical windows can overstate robustness.
• Live data, liquidity, financing, order timing, and implementation can differ from modeled assumptions.
• Models and market relationships can degrade or fail under new regimes.