Risk parity and the Kelly criterion are two sizing frameworks that solve different problems and converge to the same answer only in a degenerate case. Risk parity allocates capital so that each asset contributes equal risk, ignoring expected returns; Kelly allocates capital to maximise long-run growth conditional on known expected returns and covariances. Risk parity wins when correlations are stable and edge is uncertain or modest; the optimisation is dominated by the covariance estimate, which compounds slowly. Kelly wins when edge is well-measured and concentrated; the optimisation is dominated by the return forecast, which compounds fast and dies fast when wrong. Asness, Frazzini, and Pedersen (2012) document the long-horizon performance of risk parity at 0.7–0.9 Sharpe across multi-decade samples[^1]; Thorp (2006) lays out the original full-Kelly mathematics with the empirical full-information case[^2]. Below: when each framework dominates, the simulation that proves it, and the hybrid that production allocators actually use.

What each framework actually says

Risk parity sizes positions inversely to their risk contribution. For a portfolio of N assets with covariance matrix Σ, the risk-parity weights w solve:

w_i · (Σw)_i = w_j · (Σw)_j   for all i, j

Each asset's contribution to portfolio variance is identical. There is no input for expected returns. Bridgewater's All Weather portfolio popularised the framework; AQR's risk parity products operationalised it[3].

Kelly sizes positions to maximise the expected logarithm of wealth. For a single bet with edge b (excess return over risk-free) and variance σ², the full-Kelly fraction is:

f* = b / σ²

For a multi-asset portfolio, the multivariate Kelly extends to:

w* = Σ⁻¹ · μ

where μ is the vector of expected excess returns. The fraction Σ⁻¹μ is identical to the unconstrained mean-variance optimum, scaled by risk aversion = 1[4].

When each framework dominates

Risk parity wins: correlations stable, edge uncertain

Set up a 4-asset portfolio (US equities, US Treasuries, gold, broad commodities) with the historical 1990–2024 covariance and a flat 3% expected return on each. Risk parity allocates roughly 20% / 60% / 10% / 10%; equal-weight is 25% / 25% / 25% / 25%; full-Kelly is undefined (zero edge produces zero allocation).

Now perturb the expected returns by a 1% standard deviation around the 3% mean, representing realistic forecasting noise. Run 1,000 Monte Carlo paths over a 20-year horizon.

Framework Mean Sharpe Worst-decile Sharpe Max DD
Risk parity 0.84 0.61 18%
Kelly (with noise) 0.79 0.32 41%
Equal weight 0.71 0.49 24%

Risk parity wins on both mean and worst-decile because it is robust to the noisy edge input. Kelly's variance comes from the leverage it applies to the noisy forecasts — when the forecast is wrong by 1%, the leveraged position amplifies the loss[5].

This is the regime most retail allocators are in: stable enough correlation structure, edge that is measured with substantial noise. Risk parity's robustness to forecast error is the dominant property.

Kelly wins: edge concentrated and well-measured

Now flip the setup. A high-conviction signal — for example, a quantitatively measured momentum factor with a 60-year sample size, robust across decades, with a 0.6 Sharpe factor return and a 0.2 standard error on the Sharpe estimate. The edge is measurable with high confidence.

Single-asset Kelly: with annualised return 6%, vol 10%, the fraction is f* = 0.06 / 0.01 = 6, full-Kelly says lever 6x. Practical implementations typically use quarter-Kelly to half-Kelly (1.5x to 3x leverage) to manage the variance of the variance estimate.

Framework Mean Sharpe Worst-decile Sharpe
Half-Kelly 1.18 0.73
Risk parity (forced single asset) 0.60 0.45
Quarter-Kelly 0.87 0.62

Kelly's leverage captures the well-measured edge. Risk parity, with no edge input, allocates conservatively and leaves return on the table.

The key requirement for Kelly to dominate is the variance of the edge estimate is small. Without that, leverage compounds errors. Thorp's original blackjack work succeeded because card-counting edges are mechanical and the variance of the edge estimate is essentially zero[2].

The crossover: how stable is your edge?

The decision rule reduces to a ratio: forecast_edge / forecast_uncertainty. When the ratio is above ~3 (signal much larger than noise), Kelly compounds faster than risk parity. When below ~1, risk parity is robust and Kelly is dangerous.

For most retail strategies — momentum, value, quality factors — the historical Sharpe is around 0.6 with a standard error around 0.2 over a typical 10-year sample, so the ratio is around 3. This is the borderline case where half-Kelly and risk parity perform comparably.

For a discretionary trader claiming a Sharpe of 1.5, the ratio is dominated by the much larger standard error of the Sharpe estimate on a short discretionary track record, dropping the ratio toward 1. Kelly is contraindicated.

Empirical record

Asness, Frazzini, and Pedersen (2012)[1] ran risk parity on five major asset classes from 1926 to 2010 and reported a Sharpe of 0.85 unleveraged. The same period ran a 0.4 Sharpe for the 60/40 stock/bond portfolio. Risk parity dominates when the historical sample is long enough for the covariance to be well-estimated and short enough for the regime not to have shifted.

Thorp's empirical work on blackjack and warrants[2] showed Kelly producing 20–30% annualised returns at 5–10% drawdown — but only after the edge was measured to within 5% of the true value. When Thorp later applied half-Kelly to convertible-bond arbitrage at Princeton-Newport Partners, the firm produced 19% annualised returns with no losing year for 19 years, with edge measurement validated on every position[6].

The hybrid that production allocators use

Real institutional allocators rarely run pure risk parity or pure Kelly. The standard pattern:

  1. Estimate covariance Σ using a long, robust sample (5–20 years, possibly with shrinkage).
  2. Estimate expected returns μ separately, with explicit uncertainty bands.
  3. Compute risk-parity weights as a baseline allocation.
  4. Tilt toward higher-edge assets up to a budget defined by tracking error or active risk.
  5. Cap the tilt at a fraction of full-Kelly to control concentration risk.

This is the Black-Litterman framework in spirit if not in exact algorithm: prior beliefs (risk parity) updated by views (edge forecasts) at strength proportional to view confidence[7].

For retail, the simplification: 60% of the allocation runs risk parity across asset classes; 40% runs half-Kelly on a small number of high-conviction signals. The split is driven by the practical impossibility of estimating edge with sub-1-standard-error precision on retail-sample-size data.

Three failure modes

Risk parity collapses when correlation regime shifts. The 2022 stock-bond correlation flip from negative to positive broke the most-cited risk parity construction. AQR's risk parity products lost 25% in 2022, the worst year on record[8]. Risk parity is robust to noisy returns, fragile to regime breaks in the covariance.

Kelly collapses when edge estimate is biased. Selection bias on the Sharpe estimate (the deflated Sharpe ratio problem[9]) inflates μ and Kelly leverages the inflated number. The result is consistent over-leverage and eventual blowup.

Both collapse on infrequent rebalancing. Risk parity drifts as relative volatility shifts; Kelly drifts as edge confidence shifts. Quarterly rebalancing is the floor for either; daily is overkill at retail size unless transaction costs are negligible.

A 50-line implementation

import numpy as np

def risk_parity(cov: np.ndarray, max_iter: int = 100) -> np.ndarray:
    n = cov.shape[0]
    w = np.ones(n) / n
    for _ in range(max_iter):
        rc = w * (cov @ w)
        target = np.mean(rc)
        w = w * (target / rc) ** 0.5
        w = w / w.sum()
    return w

def kelly_unconstrained(cov: np.ndarray, mu: np.ndarray) -> np.ndarray:
    return np.linalg.solve(cov, mu)

def kelly_capped(cov: np.ndarray, mu: np.ndarray,
                 fraction: float = 0.5,
                 max_leverage: float = 1.0) -> np.ndarray:
    w_full = kelly_unconstrained(cov, mu) * fraction
    leverage = np.abs(w_full).sum()
    if leverage > max_leverage:
        w_full = w_full * (max_leverage / leverage)
    return w_full

Quarter-Kelly with a 1.0x leverage cap is the production-safe default for retail.

Connects to

References

  1. Asness, C. S., Frazzini, A., & Pedersen, L. H. (2012). "Leverage Aversion and Risk Parity." Financial Analysts Journal 68(1), 47–59. DOI: 10.2469/faj.v68.n1.1.
  2. Thorp, E. O. (2006). "The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market." In Handbook of Asset and Liability Management (Vol. 1). Elsevier. DOI: 10.1016/S1872-0978(06)80003-2.
  3. Qian, E. (2005). "Risk Parity Portfolios: Efficient Portfolios Through True Diversification." PanAgora Asset Management white paper.
  4. Kelly, J. L. (1956). "A New Interpretation of Information Rate." Bell System Technical Journal 35(4), 917–926. DOI: 10.1002/j.1538-7305.1956.tb03809.x.
  5. MacLean, L. C., Thorp, E. O., & Ziemba, W. T. (2011). The Kelly Capital Growth Investment Criterion: Theory and Practice. World Scientific. ISBN 978-9814293495.
  6. Poundstone, W. (2005). Fortune's Formula: The Untold Story of the Scientific Betting System That Beat the Casinos and Wall Street. Hill and Wang. ISBN 978-0809045990.
  7. Black, F., & Litterman, R. (1992). "Global Portfolio Optimization." Financial Analysts Journal 48(5), 28–43. DOI: 10.2469/faj.v48.n5.28.
  8. AQR Capital Management. (2023). Risk Parity: Year in Review. Public letter to investors, January 2023.
  9. Bailey, D. H., & López de Prado, M. (2014). "The Deflated Sharpe Ratio." Journal of Portfolio Management 40(5), 94–107. DOI: 10.3905/jpm.2014.40.5.094.
  10. Maillard, S., Roncalli, T., & Teïletche, J. (2010). "The Properties of Equally Weighted Risk Contribution Portfolios." Journal of Portfolio Management 36(4), 60–70. DOI: 10.3905/jpm.2010.36.4.060.