A regime-aware systematic equity strategy engineered to generate consistent alpha through ensemble machine learning — the institutional rebuild of ProjectV, and 00Capital's flagship.
Real-time paper-trading results, published unedited — the research and architecture behind them follow below. Holdings are disclosed as portfolio weights.
Adapt to the market you're actually in
Static strategies die in regime shifts. A momentum system tuned for trending markets bleeds in chop; a mean-reversion system gets steamrolled in a crisis. ProjectR's core premise is that positioning should be conditional on the market state — identified explicitly, not assumed.
A six-state regime classifier sits at the top of the stack. Beneath it, an ensemble of XGBoost and LightGBM gradient-boosting models — trained on 20+ years of historical data — votes on positioning. The ensemble structure reduces overfitting and improves generalization to market conditions the models have never seen.
ProjectR exists because its predecessor failed honestly: ProjectV revealed architectural flaws under stress testing, and rather than patch them, we rebuilt the system to an institutional standard from the ground up.
Backtested results are hypothetical, do not reflect live trading, and have inherent limitations — including the benefit of hindsight and the absence of real slippage and liquidity constraints. They are research output, not a promise. Stress scenarios covered: 2008 financial crisis, 2020 COVID crash, 2022 rate-hike cycle, and 1987 Black Monday.
Six market states. One adaptive system.
The market is not one environment — it is several, and they punish the wrong posture. The classifier identifies the active state and the decision engine repositions accordingly, instead of betting one static style survives them all.
Regime classification layer
A six-state classifier identifies whether the market is trending, mean-reverting, volatile, in crisis, or flat — the conditioning variable for everything downstream.
Ensemble decision engine
XGBoost and LightGBM models trained on 20+ years of data vote on positioning. Disagreement between models is itself a signal — conviction scales exposure.
Risk-controlled construction
Position sizing and portfolio construction enforce strict drawdown discipline, targeting sustainable outperformance rather than headline returns.
Live paper execution via Alpaca
The strategy trades a paper account through the Alpaca API, with the trading service hosted on Railway and deployed straight from Git — the equity curve on this page reads from that account.
Out-of-sample testing
Held-out years the models never trained on. If performance only exists in-sample, it isn't performance.
Walk-forward optimization
Parameters are re-fit only on data available at each point in time, eliminating look-ahead bias from the results.
Monte Carlo permutation tests
Returns are tested against randomized orderings to verify the edge is statistically distinguishable from chance.
Crisis stress replay
2008, 2020, 2022, and 1987 scenarios replayed through the full system — the strategy must survive its worst decade, not just its best.