What is walk-forward testing?
The gold standard
A standard backtest trains a model on historical data and then tests it on that same data. This is like grading an exam using the answer key the student has already seen. The results look good, but they prove nothing about real-world performance.
Walk-forward testing is different. The model is trained on a window of past data, then tested on the next period of genuinely future data it has never encountered. The window rolls forward through time. At every step, the model faces data it could not have been optimized for.
This is how we validate every strategy before it enters the portfolio. If it cannot perform on data it has never seen, it does not trade.
The results below span three measurement windows. Each tells a different part of the story, and each comes with honest caveats about what it does and does not prove.
Portfolio results
Crisis performance
A system's character is revealed by what it does when everything else breaks. Two events define the recent era of market stress. The portfolio was positive in both.
These are walk-forward results from the strategies that were active during each period. The system rotated into protective positions not because it predicted the crash, but because its regime detection identified the shift in real time and the portfolio's multi-engine structure naturally hedges directional exposure.
What we don't hide
Transparency is not a section we add because it looks good. It is the reason this document exists. If you cannot verify a claim, it is not a claim worth making. Here is what you need to know before interpreting the numbers above.
Honest caveats
The recent period was mostly bull market. The 3.5-year overlap window (Nov 2022 - Apr 2026) is 86% bull regime. This flatters every equity-adjacent strategy. The 3.31 Sharpe reflects favorable conditions, not a guarantee.
Harvest's returns are inflated by mega-cap concentration. Harvest's 46% CAGR is driven by concentrated positions in AAPL, MSFT, GOOG, and AMZN during the strongest tech run in market history. Our forward estimate for Harvest alone is 25-30% CAGR. The number will normalize.
The real max drawdown is larger than the observed one. The bull-period observed max drawdown was -9.8%. Our stress-estimated max drawdown, which accounts for correlated selloffs across strategy legs, is -17.5%. We report the stress number because it is more honest.
These are walk-forward results, not live audited returns. We are pre-launch. Walk-forward testing is the most rigorous form of backtesting, but it is not the same as live trading with real capital, real slippage, and real market impact. We will provide audited results once live trading begins.