The 2nd-place capstone — the optimizer. Clever, clears the hurdle… and still can't beat humble risk-parity. DeMiguel's 1/N, on the family's own keepers.
Brilliant is one of the three capstone allocators of the Blaque Baux family (the runner-up to breakthrough). Same keeper set, same engine — a smarter allocation rule.
Not investment advice. Educational/research software. Nothing here is validated to a live-money bar. See LICENSE.
git clone --recursive https://github.com/blaquebaux/brilliant.git
julia --project=engine -e 'using Pkg; Pkg.instantiate()' # one-time engine setupThere are three ways to assemble the keepers, and they embody three philosophies: humility (breakthrough — risk parity, ignores return forecasts), cleverness (brilliant — optimize using estimated returns and covariance), and aggression (bossy — concentrate on the winners). Brilliant is the clever one: mean-variance max-Sharpe (tangency) and min-variance over the six keepers. The honest question is the oldest one in portfolio construction — does optimization survive estimation error out-of-sample, or does the humble baseline win? (DeMiguel, Garlappi & Uppal: 1/N is hard to beat.)
research/brilliant_1_optimized.py — same keepers as breakthrough,
causal, monthly, net of cost, long-only (Alpaca SIP 2016–2026):
| allocator | Sharpe | CAGR | vol | maxDD | M² exc |
|---|---|---|---|---|---|
| risk-parity (breakthrough) | +1.16 | +11% | 9.3% | −17% | +5.1% |
| max-Sharpe (mean-var) | +0.90 | +10% | 11.8% | −23% | +0.5% |
| min-variance | +1.09 | +8% | 7.1% | −15% | +3.8% |
| SPY (Bogle hurdle) | +0.87 | +15% | 18.0% | −34% | — |
Optimization does not beat humble risk-parity. The "brilliant" max-Sharpe optimizer, which uses estimated returns, overfits to +0.90 — below the humble baseline and barely above the index — exactly the estimation-error trap DeMiguel warned of. Min-variance (which uses only covariance, not returns) is the honest close second at +1.09, with the lowest drawdown of any allocator (−15%) — a legitimate alternative for the drawdown-averse. But nothing clever beats the return-agnostic +1.16.
So brilliant is a keeper-tier 2nd-place capstone: it clears the Bogle hurdle, and min-variance is a defensible choice — but it's as good, not better, and more fragile (the more it uses forecasts, the worse it does). The family's ethos, proven at the allocation layer: humility beats cleverness.
Research validated — the 2nd-place capstone. Optimizing the keeper allocation (max-Sharpe, min-variance) clears the index but does not beat breakthrough's humble risk parity; the return-using optimizer overfits (+0.90), the covariance-only min-variance ties-but-doesn't-beat (+1.09). DeMiguel's 1/N, on the family's own keepers. A legitimate alternative allocator, ranked 2nd. The governed version rides on breakthrough's allocator.
Blaque Baux is a quantitative research initiative and a subsidiary of Carter Warrens. BlaqueBaux.com is the home for the work; the code lives here on GitHub — open to study, test, and build bespoke strategies on top of.
Anyone can point an AI at a market. The edge is understanding what the data actually says — and turning it into something you can act on. We test relentlessly and put most of it on the record as rejected, with the reason; what survives is built, governed, and validated before it is ever called real. That combination — honest research, reproducible evidence, and execution you can trust — is why Carter Warrens leads on strategy and implementation, not merely uses the tools everyone now has.
This repo is one sleeve of the Blaque Baux family — a single governed engine steered in many directions. The core repo is the base/blueprint and holds the full family roster.
engine/ the Blaque Baux platform (git submodule -> blaquebaux/base; incl. PortfolioOpt)
research/ _brilliant_common.py + brilliant_1_optimized.py (the optimizer bake-off) + scorecard
live/ (rides on breakthrough's governed allocator)
MIT. (c) 2026 Carter Warrens.
This capstone is wired to its own governed rail — the min-variance optimizer. live/brilliant_allocator.py emits today's
target book (as of the last settled close); live/brilliant_live.jl routes it through the engine's Layer-3 safety
gate (preflight, idempotency, reconciliation, HWM, kill switch) — no LLM in the order path. Dry-run by
default (live/run_brilliant_daily.sh), graduates to paper once ~/.config/blaquebaux/alpaca_brilliant.env exists.
Capstones are alternatives, not additions — stacking them is net-negative (see breakthrough's
capstone_of_capstones); run the one whose objective fits. Dry-run PASSES the gate.