Testing a volatility-conditioned position budget

Our historical fleet record found median chosen leverage at 5.0x across volatility sextiles spanning a 5.7x volatility spread. Losses and liquidations worsened in the more volatile groups. That observation suggests a practical question: can a position budget respond to volatility while preserving the agent's trade selection? We describe a PROPOSED hypothesis test, rather than a measured DXAP improvement or a promised product setting.

Change exposure, preserve the proposed trade

The reference arm uses the saved proposed size. The treatment applies a documented volatility-conditioned notional cap to the same proposed symbol, direction and decision time. A simple candidate formula is budget multiplied by reference volatility divided by observed volatility, bounded above by the owner's existing exposure limit. All inputs must precede the order. We would define the volatility estimator, observation interval and minimum history before inspecting outcomes.

Consider an illustrative $1,000 reference notional at 2% volatility. At 4% observed volatility, the formula yields a $500 cap. At 1%, it yields $2,000 before the owner's maximum cap is applied. The arithmetic is straightforward; the hard question is whether the volatility estimate captures the risks that cause losses. A jump immediately after a quiet window can defeat a backward-looking measure.

Measure what the cap actually changes

We would report exposure, drawdown, liquidation incidence and missed positive outcomes. Smaller positions can reduce dollar losses mechanically while leaving return per unit of exposure unchanged. That is useful risk control, but it should be described accurately. Comparing only total P&L would obscure whether the cap improves trade quality or simply runs less capital through the same decisions.

The treatment must also specify leverage and margin behavior. Equal notional under different leverage can produce different liquidation distances and different capital requirements. The historical paper engine used zero slippage and zero funding with a maintenance-margin placeholder that differs from live venue margining. Reusing those fills would support an initial replay, while a live-execution claim would require venue-faithful liquidation and cost handling.

Search discipline matters as much as the formula

We would choose a small, declared set of estimator windows and cap strengths, fit on an earlier period and freeze a choice for a later holdout. Publishing only the best setting from a large sweep would exaggerate the result. Results should include losses clustered by market day and a constant-exposure comparison, so a risk reduction can be assessed against the simpler alternative of reducing every order equally.

The source findings and their scope are available in the continuous record; the controls paper explains the distinction between model proposals and deterministic policy. DXAP publicly describes limits checked outside the model. That provides a concrete architectural place to study a cap. Whether this particular formula deserves implementation depends on the controlled comparison, the owner's intended risk budget and evidence that it improves the relevant failure mode.