Would this strategy actually have made money?
A shared backtesting harness for the Minutemen Alternative Investment Fund. Every member's strategy runs through the same pipeline — same data, same metrics, same stress tests — so results are directly comparable instead of each person producing a different one-off notebook that can't be graded against anyone else's.
Three strategies, one window.
Representative equity curves across 2022–2025 on SPY. The covered-call writer outperforms the underlying by nearly 40 percentage points; a mechanical SMA crossover roughly matches buy-and-hold after frictions.
Intermediate shape is synthesized from the published endpoint returns — actual
per-day series ship once export_scorecard_json() lands.
Survive more than one way of being wrong.
- § 01
Monte Carlo stress
Six synthetic-data generators — GBM, block bootstrap, regime switching, noise injection, Heston stochastic vol, and the trained cGAN — run automatically. If a strategy shows positive Sharpe on pure GBM noise, it is overfitting.
280+ backtests · per scorecard - § 02
GAN regime scenarios
A conditional Wasserstein GAN trained on four real SPY regimes — bullish, bearish, sideways, crash — generates unlimited synthetic paths that preserve regime-specific volatility clustering and tail behavior.
- § 03
Engine divergence
Every strategy runs through both a bar-based engine (market orders at next open) and an event-driven engine (stop / limit / OCO fills intrabar). If the engines disagree, the strategy's edge is sensitive to execution — a flag that live results will differ.
“The disagreement is the point. A strategy that only works under one engine, one asset, or one historical window isn't a strategy — it's a fit.”
Write a strategy. Get a scorecard.
Three lines of user code, one 4-page scorecard PNG, letter grades on every dimension. Members bring their own logic; the framework handles data, execution, frictions, stress tests, and reporting.