Objective
Continue issue 1310 with a target above 30% full-window CAGR, maximum drawdown no worse than -25%, and only UPRO/cash, while explicitly allowing aggressive hindsight fitting and requiring drawdown-episode indicator analysis.
Results
- Evaluated 1,873 deterministic trials: 1,816 causal full-window fits and 57 memorized-event or forward-return hindsight variants.
- Sixteen variants clear both numerical constraints, all of them non-causal.
- Winner
l3x.hindsight.036f423a11de: 1,942.14% CAGR, -0.14% maximum drawdown, UPRO/cash only, but it knows the same-session UPRO return before choosing exposure and cannot be traded. - Zero causal variants clear both constraints. The issue 1310 baseline remains the best feasible causal result at 8.32% CAGR and -24.78% drawdown.
- The highest-CAGR causal veto reaches 17.28% CAGR but fails with -50.63% drawdown.
Drawdown analysis
The eight largest distinct parent-baseline drawdowns required different defenses. Lagged macro led during dot-com, GFC, and July 2024; fast 20-day trend led in 2018 and COVID; 20-day realized volatility led in 2022; breadth led in 2015–16; and the full ensemble led in 2011. The evidence does not support one causal crash indicator that avoids every episode.
Artifacts
- Implementation:
research/leveraged_sp500_hindsight.py - Ledger and diagnostics:
research/campaigns/leveraged-sp500-indicator-ensemble/hindsight-trials.json,hindsight-diagnostics.json - Executed notebook:
research/experiments/1313.ipynb - Public analysis:
docs/experiments/1313-analysis.html - Interactive research page:
docs/research/leveraged-sp500-indicator-ensemble.html
Limitations
The numerical success is a deliberately non-causal research ceiling, not a strategy or promotion candidate. Causal candidates are still selected in sample; pre-2009 UPRO is synthetic; breadth and macro inputs have documented proxy/vintage limitations; and transaction costs, taxes, spreads, cash yield, and intraday execution are simplified.