Methodology notes — firm-wide
The question that matters
We build and validate systematic crypto strategies across live and forward paper stages, and we do not manage client assets. The question we care about most is not whether a backtest looks good — it's how much of that backtest will survive contact with live markets. This note walks through the validation stack we run before any strategy graduates toward live capital. It's deliberately concept-level: the how, not the proprietary parameters.
Walk-forward, not a single fit
One train/test split on a single window tells you almost nothing about robustness — it tells you a configuration could be fit to that window. We use walk-forward validation instead: re-fit on a trailing window, evaluate only on the immediately following out-of-sample segment, then roll the window forward and repeat. Every result we act on is out-of-sample by construction. Anything that only works when it has effectively seen the future gets filtered out early. On the AI-signal side we use an ensemble rather than a single model, but the discipline is identical — it's judged only on data it was never fit on.
Worst-case intra-bar fill reconstruction
The most common way a backtest quietly lies is in its fill assumption. Fill at the close, or assume you always caught the mid, and you're borrowing performance from the future. So we reconstruct each bar's path in the least-favorable order — for an up bar, open → high → low → close; for a down bar, open → low → high → close — and assume our entries and stops trigger along that adverse path, with slippage and fees charged on top. The goal isn't pessimism for its own sake; it's to make sure the edge that survives is the part that doesn't depend on optimistic execution.
We learned the value of this the practical way. When we began tracking live execution against the model, our effective per-trade cost came in somewhat above the flat assumption we'd been using — mostly maker orders timing out and converting. The right response was to revisit the cost assumption in validation, not to start tinkering with the live system. Worst-case fills are what give you the headroom to absorb that kind of surprise instead of getting blindsided by it.
Regime awareness
Crypto doesn't sit still. A carry/basis environment and a trending environment reward completely different behavior, and a system that's unconditionally "on" tends to give back its gains when the regime flips. We treat a regime-transition signal as information about when an edge is likely to be present — not as a price prediction. In practice that means some strategies are deliberately flat for stretches. Being out of the market is a position.
Drift monitoring doesn't stop at deployment
Validation isn't a one-time gate. We run a daily check comparing live behavior against the backtest distribution across a set of residual diagnostics, with a change-point test layered on top so a slow drift raises a flag before it becomes a drawdown. When live and backtest diverge, we treat it as a research trigger — not something to explain away.
On performance
You'll notice there are no return figures anywhere in this note, and that's deliberate. Our policy is simple: we publish the methodology, never the performance numbers. Returns, win rates, and equity curves are shared individually with qualified investors on request. The methodology is the part you can actually evaluate from the outside — so that's what we put in public.
If you build validation stacks and think we've got something wrong, we'd genuinely like to hear it.
Disclaimer: This is a research note for informational purposes only. It is not investment advice, a solicitation, or an offer to buy or sell any asset. Performance data is shared individually with qualified investors on request.