Return Dynamics & Position Sizing

Introduction

Position sizing is conventionally framed through risk tolerance. A complementary framing comes from the return process itself: the same capital can produce materially different geometric outcomes depending on whether the underlying is trending, oscillating, or directionless.

The argument generalises across single names, indices, asset classes, and systematic strategies. Leverage does not change the mechanics; it amplifies the differences and so makes the regime sensitivity easier to observe.

What follows traces the mathematics behind that framing and the empirical support from recent leveraged ETF research.

The Compounding Problem

Returns compound, and compounding is asymmetric: a 50% loss requires a 100% gain to recover. Long-run wealth therefore depends on how position size interacts with the return path, not just on average returns.

Two distinct effects fall out of this asymmetry. They are often conflated in practice, but the mechanics behind them differ.

Compounding asymmetry (volatility drag)

Even with no borrowing, oscillating returns erode wealth:

$+15%$ then $-15%$ = $0.9775$ (a loss of 2.25%, not zero)
With 2 $\times$ position size: $+30%$ then $-30%$ = $0.91$ (a loss of 9%, not 4.5%)

The excess loss beyond the 2 $\times$ scaling is pure compounding asymmetry — a mathematical property of multiplicative returns, not a function of fees, costs, or rebalancing.

What governs whether scaling up a position helps or hurts is not the absolute level of drift, but drift relative to variance. A high-drift, high-variance asset can carry a low or even negative $\mu/\sigma^2$ , in which case additional leverage compounds against the holder. A low-drift, low-variance asset can carry a high $\mu/\sigma^2$ , in which case it does not.

The relevant quantity is the ratio. The numerator and denominator individually carry less information than their combination.

Rebalancing drag

A second drag arises only when position size is mechanically restored to a fixed ratio each period — as leveraged ETFs do daily. Rebalancing forces buying after gains and selling after losses, producing a buy-high-sell-low cycle in oscillating markets.

In trending markets the same mechanism becomes a tailwind: it adds exposure as momentum builds.

Source of drag	Exists without rebalancing?	Direction in trending market
Compounding asymmetry	Yes — always present	Reverses — becomes a tailwind
Rebalancing drag	No — only with rebalancing	Reverses — adds momentum

Market regime and outcome

The same position size that compounds favourably in a trending market can compound against the holder in a mean-reverting one. The magnitude of this effect is large relative to the dispersion in average returns across regimes.

Regime	Return pattern	Larger position vs smaller	Why
Trending	Persistent direction	Larger wins more than proportionally	Compounding amplifies each successive gain on a growing base
Sideways	No clear direction	Similar, slight drag	Compounding asymmetry creates small losses; rebalancing is neutral
Mean-reverting	Up then down repeatedly	Larger loses more than proportionally	Compounding asymmetry destroys wealth; rebalancing makes it worse

Hsieh, Chang and Chen (2025) confirm this regime-dependence empirically across 20 years of S&P 500 and Nasdaq-100 data. Rolling autocorrelation tracks realised compounding effects: positive autocorrelation produces consistent outperformance, negative autocorrelation consistent underperformance.

The Key Signal: $\mu/\sigma^2$

Position size, expressed mathematically, reduces to a single ratio. The geometric (long-run compounded) return has an exact continuous-time form:

Geometric return	$G = \mu - \tfrac12\sigma^2$	Exact in continuous time (Itô's lemma). In discrete time this is a close approximation for small periodic returns.

Scaling the position by factor $\beta$ ( $\beta = 2$ means double the base position) turns each period's return into $\beta \times r$ , so:

Scaled position	$G_\beta = \beta\mu - \tfrac12\beta^2\sigma^2$	Mean scales linearly with $\beta$ ; variance penalty scales with $\beta^2$ .

The mean term scales linearly with $\beta$ , but the variance penalty scales with $\beta^2$ . Doubling position size doubles expected gain and quadruples the variance penalty.

The relative weight of these two terms is what makes the ratio $\mu/\sigma^2$ the dominant variable. When the ratio is high the linear gain term is ahead; when it is low the quadratic penalty is ahead.

General break-even condition

A scaled position with leverage $\beta$ beats the base position ( $\beta = 1$ ) when $G_\beta > G_1$ . Solving for any $\beta > 1$ :

Break-even ( $\beta > 1$ )	$\mu / \sigma^2 > \tfrac12(\beta + 1)$	Must hold for the larger position to outperform the base position.

For inverse or short positions ( $\beta < 0$ ), the inequality flips:

Break-even ( $\beta < 0$ )	$\mu / \sigma^2 < \tfrac12(\beta + 1)$	Market must have sufficiently negative drift for a short position to win.

$\mu/\sigma^2$ is the signal-to-noise ratio of returns: drift relative to variance. The threshold each leverage level must clear:

Position size $\beta$	Break-even: $\mu/\sigma^2$ must exceed	Threshold value	Interpretation
1.5 $\times$	$\tfrac12(1.5 + 1) = \tfrac12(2.5)$	$> 1.25$	Mild trend sufficient
2 $\times$	$\tfrac12(2 + 1) = \tfrac12(3)$	$> 1.50$	Moderate trend required
3 $\times$	$\tfrac12(3 + 1) = \tfrac12(4)$	$> 2.00$	Strong trend required
4 $\times$	$\tfrac12(4 + 1) = \tfrac12(5)$	$> 2.50$	Very strong trend required
$-1\times$ (short)	$\tfrac12(-1 + 1) = \tfrac12(0)$	$< 0.00$	Negative drift needed
$-2\times$	$\tfrac12(-2 + 1) = \tfrac12(-1)$	$< -0.50$	Strong negative drift
$-3\times$	$\tfrac12(-3 + 1) = \tfrac12(-2)$	$< -1.00$	Very strong negative drift

Each additional unit of leverage raises the required signal-to-noise ratio by 0.5. Going from 2 $\times$ to 3 $\times$ does not require slightly more momentum — it requires the ratio to be 33% higher.

The Kelly criterion: optimal position size

Setting the derivative of $G_\beta$ with respect to $\beta$ to zero gives the position size that maximises long-run geometric return:

Kelly optimal size	*$f^ = \mu / \sigma^2$**	Under these assumptions, the signal-to-noise ratio is the geometric-return-maximising position size.

Under these assumptions the geometric-return-maximising size equals $\mu/\sigma^2$ : a ratio of 1.5 corresponds to 1.5 $\times$ the base position, 0.3 to a 30% position, and a negative ratio to a short. The Kelly value is the long-run optimum on the assumed return process, not the recommended live size. In practice, fractional Kelly is the more defensible choice, since estimates of $\mu$ and $\sigma$ carry meaningful error and full Kelly is highly sensitive to that error.

Plugging $f^* = \beta$ back into the break-even condition gives $\mu/\sigma^2 > \tfrac12(f^* + 1)$ , which holds whenever $f^* > 1$ . Kelly leverage outperforms the unleveraged position when the signal-to-noise ratio exceeds 1.

*$f^ = \mu/\sigma^2$**	Signal strength	Position sizing action	Typical regime
$> 2.0$	Very strong	2 $\times$ or more	Strong momentum, bull trend
$1.0 - 2.0$	Moderate	Scale between 1 $\times$ and 2 $\times$	Mild trend, recovering market
$0.5 - 1.0$	Weak	Below full allocation	Low-signal, near-random
$\approx 0$	None	Minimal position or cash	Sideways, no clear drift
$< 0$	Negative	Short or exit entirely	Mean-reverting or bear market

Rebalancing Frequency

Given a target position size, the rebalancing frequency that minimises drag is a function of return autocorrelation rather than the calendar.

Autocorrelation	Market character	Rebalance frequency	Logic
Positive ( $\varphi > 0$ )	Momentum / trending	Daily — as frequent as possible	Restoring size adds exposure in the direction of the trend
Near zero ( $\varphi \approx 0$ )	Independent / random	Frequency does not matter	No autocorrelation means no rebalancing benefit or cost
Negative ( $\varphi < 0$ )	Mean-reverting / oscillating	Weekly or monthly — reduce frequency	Less frequent rebalancing gives the market less opportunity to whipsaw the position

Hsieh et al. (2025) show that shifting from daily to monthly rebalancing nearly eliminates drag in mean-reverting regimes while costing little in trending ones.

Reference: Hsieh, Chang & Chen (2025). Compounding Effects in Leveraged ETFs: Beyond the Volatility Drag Paradigm.