Adaptive position sizing: scale exposure with a signal
A fixed bet ignores the weather. Adaptive position sizing scales how much you trade with a live signal — volatility, the VIX, trend strength — so you press when conditions favour you and pull back when they don't. The catch: it's a powerful new way to overfit. Here's how to do it, and how to prove the curve holds out-of-sample.
What is adaptive position sizing?
Adaptive (or dynamic) position sizing makes your trade size a function of a driver series instead of a constant. The classic case is volatility targeting — size inversely to volatility so every trade risks about the same. But the driver can be anything: the VIX, an index's trend strength (ADX), an indicator, or another asset. You map the driver through a response curve to an exposure multiplier — for example VIX 15 → 1.5×, VIX 35 → 0× (risk-off).
The four response-curve shapes
- Linear — a straight line between two points. Two knobs, easy to read, hard to overfit.
- Sigmoid — a smooth S-curve with a midpoint and steepness; ramps, then saturates at the extremes.
- Exponential — a saturating curve for an aggressive early response.
- Piecewise / step — discrete regimes ("VIX < 15 → 1.5×, 15–25 → 1×, > 30 → 0×"). The most flexible — and the easiest to curve-fit.
Common adaptive-sizing recipes
- VIX risk-off — cut size, or stop entering, as implied volatility rises.
- Inverse-volatility / vol targeting — size down when realized volatility (ATR) is high.
- Trend scaling — press size when trend strength (ADX) is high, ease off in the chop.
- Regime gate — block new entries entirely outside a favourable regime.
Why it usually overfits — and the test that catches it
Every knob you add to a response curve is a degree of freedom the optimizer can bend to fit the past. A free-form step function with five tunable breakpoints can trace the noise in your sample and look spectacular in a backtest — then do nothing live. Serious practitioners have avoided building configurable-curve sizing tools precisely for this reason.
The honest test is whether the relationship the curve exploits holds out-of-sample. Bin the driver, measure the average forward return in each bin on your in-sample window, then on a held-out window. If the in-sample and out-of-sample response curves agree, the edge is structural. If the out-of-sample curve goes flat or inverts, your curve was fitting noise.
How QUANTHEON Lab does it
The Adaptive Exposure engine makes this a first-class, visual object: pick any series, choose a curve shape, watch the mapping update live, and point it at position size, pyramiding or an entry gate. The curve's parameters run through the same anti-overfitting gauntlet as everything else — they count toward the Deflated-Sharpe haircut — and a one-click In-Sample-vs-Out-of-Sample robustness check shows whether the shape survives. You can even describe it to the AI ("cut size when VIX is high") to build it or edit your current strategy. No code.
FAQ
What is adaptive position sizing?
It scales trade size with a live signal — volatility, the VIX, trend strength — instead of a fixed bet, mapping the driver through a response curve to an exposure multiplier so you risk more in favourable conditions and less in adverse ones.
Is volatility targeting the same thing?
Volatility targeting is one special case — sizing inversely to realized volatility. Adaptive position sizing generalizes it to any driver series and any curve shape.
Does adaptive position sizing cause overfitting?
It can: each tunable point on the curve is a degree of freedom. Mitigate it by preferring few, smooth knobs, counting the curve's parameters in your overfitting checks, and confirming the in-sample response curve still holds out-of-sample.
Related: What is overfitting? · Walk-Forward analysis · Position size calculator