📦 Bilbo Box — Pareto Hill-Climb

Multi-objective optimization of the Bilbo Box breakout mechanics: maximize mean P&L while minimizing max drawdown, over 104 evaluated configurations on paired zero-drift Monte Carlo panels.

Forward window 2026-10-02 to 2026-10-29 · 25 screening paths/seed 20261002 · 150 validation paths/fresh seed 20262001 · 49.0 min compute

⛔ NOT FUNDABLE

The hill-climb pick

$879mean P&L / path (screen)

11.8%mean max drawdown

52%P(profit)

2.9trades / path

Winning configuration (diffs vs the author's baseline):

Target premium %=6.0 Kill-switch DD %=15.0 Spread cap=0.03 Signal start (ET)=9 Signal end (ET)=16 Trail keep frac=0.15 ATR run mult=0.75 Strike ATR mult=0.5 DTE target=21 Time cap (days)=10

Honest read: zero-drift mechanics optimization cannot manufacture edge. A fundable verdict requires positive out-of-sample expectancy on real data. This optimization tunes mechanics under a random walk — entry/exit/sizing geometry — not predictive edge. Every configuration, including the pick, was evaluated under zero drift.

Pareto frontier

Non-dominated configurations across all 104 evaluations. No frontier point is better on both axes than any other.

hill-climb pick frontier point author baseline

Frontier table

Config (vs baseline)	Mean P&L	Median	Mean DD	P(profit)	Kill-sw	Trades
Target premium %=6.0 Kill-switch DD %=15.0 Spread cap=0.03 Signal start (ET)=9 Signal end (ET)=16 Trail keep frac=0.15 ATR run mult=0.75 Strike ATR mult=0.5 DTE target=21 Time cap (days)=10	$879	$22	11.8%	52%	40%	2.9
Target premium %=6.0 Kill-switch DD %=15.0 Spread cap=0.03 Signal start (ET)=9 Signal end (ET)=16 Trail keep frac=0.15 ATR run mult=0.75 Strike ATR mult=1.0 DTE target=35 Time cap (days)=10	$802	$18	11.2%	52%	36%	3.1

Fresh-seed validation (150 paths)

Top-8 frontier points + baseline re-run on unseen panels. If the pick only won on the screening seed, it shows up here.

Config (vs baseline)	Mean P&L	Mean DD	P(profit)	Kill-sw
Target premium %=6.0 Kill-switch DD %=15.0 Spread cap=0.03 Signal start (ET)=9 Signal end (ET)=16 Trail keep frac=0.15 ATR run mult=0.75 Strike ATR mult=0.5 DTE target=21 Time cap (days)=10	$361	13.9%	38%	39%
Target premium %=6.0 Kill-switch DD %=15.0 Spread cap=0.03 Signal start (ET)=9 Signal end (ET)=16 Trail keep frac=0.15 ATR run mult=0.75 Strike ATR mult=1.0 DTE target=35 Time cap (days)=10	$289	13.0%	37%	33%
author baseline	$105	16.2%	32%	7%

Author's published results (reference)

Milkman's live paper log as of 2026-09-24 — the real-money-shaped anchor this research is measured against:

+66.79%since $10k start (Jul 9)

41%win rate (46 closed)

1.72profit factor

-18.25%max drawdown

50/128signals taken

Source: milkmantrades.com/bilbo-paper.html daily log. Our zero-drift simulation cannot reproduce his edge — it only stress-tests mechanics.

Parameter space searched

Trend gatestrict, off, slope, band, halfsize

Target premium %2.0, 3.0, 4.0, 5.0, 6.0

Kill-switch DD %15.0, 20.0, 25.0, 30.0

Spread cap0.03, 0.04, 0.05, 0.06

Signal start (ET)9, 10

Signal end (ET)14, 15, 16

Trail keep frac0.15, 0.25, 0.35, 0.5

ATR run mult0.75, 1.0, 1.25

Strike ATR mult0.5, 0.75, 1.0

DTE target21, 28, 35

Time cap (days)10, 14, 21

Greedy Pareto walk from the author baseline: each round evaluates every single-parameter neighbor on identical panels; moves only to a Pareto-dominating neighbor (noise epsilons: $10 P&L, 0.2pp DD). Climbed 4 rounds.

Assumptions (stated, not hidden)

drift: 0 (random walk) — mechanics stress test, not edge validation
iv: constant per symbol; calibrated from Yahoo option chain or 1.3x realized-vol fallback
pricing: Black-Scholes, r=4.5%
spread: 1-4% model, fills at mid +/- quarter-spread
warmup: real Yahoo daily+hourly bars
forward: simulated
paired_panels: identical panels per candidate within a phase (fresh Random(seed))

⬇ Full run JSON (configs, all evaluations, frontier, validation)

Research artifact · engine: bilbo.engine.Engine (episode-corrected 2026-09-22) + config tuning knobs · generated 2026-10-02 12:50 ET · not investment advice