Sortino Ratio: Formula, Downside Deviation, and Python Implementation

The Sortino Ratio is a refinement of the Sharpe Ratio that addresses a long-standing criticism: volatility is not symmetric risk. A strategy that occasionally produces large gains is penalized by the Sharpe formula even though large gains are desirable. The Sortino Ratio solves this by dividing excess return only by downside deviation — volatility from returns that fall below the target rate. Trend-following strategies and long-volatility positions score materially higher on Sortino than Sharpe, accurately reflecting their asymmetric risk profile.

The Formula

For a return series R with target return T (minimum acceptable return, often the risk-free rate or 0):

The downside deviation σ_D counts only returns below T:

Three implementation choices determine correctness:

• Target return T: Use 0 for simplicity (the most common convention in strategy comparisons) or the period-equivalent risk-free rate. The choice is material — always report which T you used.
• Sample vs. population denominator: Use n−1 (sample) for unbiased estimates, consistent with how sample standard deviation is computed for the Sharpe Ratio.
• Annualization: Multiply the period Sortino by the same square-root factor used for Sharpe: ×√252 daily, ×√52 weekly, ×√12 monthly.

Annualization

The annualization logic is identical to the Sharpe Ratio. Both assume i.i.d. returns: mean scales linearly with time, variance scales linearly, and standard deviation scales with the square root. Multiplying the period Sortino by √(periods per year) gives the annualized figure:

The square-root rule applies to the ratio as a whole, not separately to the numerator and denominator. The downside deviation is a per-period figure; multiplying the ratio by √T is equivalent to annualizing the numerator by ×T and the denominator by ×√T.

Interpretation Benchmarks

When a strategy's excess return is positive, the Sortino Ratio is ≥ the Sharpe Ratio — the same positive numerator divided by the smaller downside denominator (downside deviation ≤ total standard deviation). For a strategy with normally distributed returns, σ_D ≈ σ / √2, so Sortino ≈ Sharpe × √2 ≈ 1.41 × Sharpe. Use this relationship to sanity-check both ratios:

Python Implementation

import math

def sortino_ratio(returns: list[float], risk_free_rate: float = 0.0) -> float:
    """
    Compute the period Sortino Ratio (not annualized).

    Args:
        returns:         list of period returns as decimals (e.g., 0.01 for 1%)
        risk_free_rate:  period target / minimum acceptable return (MAR)

    Returns the Sortino Ratio for the period. Multiply by an annualization
    factor to get the annualized Sortino.
    """
    n = len(returns)
    if n < 2:
        raise ValueError("Need at least 2 data points")

    mean_r = sum(returns) / n
    excess  = mean_r - risk_free_rate

    # Downside deviation: only returns below MAR contribute
    downside_sq  = sum(min(r - risk_free_rate, 0) ** 2 for r in returns)
    downside_std = math.sqrt(downside_sq / (n - 1))  # sample denominator

    if downside_std == 0:
        return float("inf") if excess > 0 else 0.0
    return excess / downside_std


ANNUALIZATION = {
    "daily":   math.sqrt(252),
    "weekly":  math.sqrt(52),
    "monthly": math.sqrt(12),
    "annual":  1.0,
}

def annualized_sortino(returns: list[float], periodicity: str = "daily",
                         risk_free_rate: float = 0.0) -> float:
    factor = ANNUALIZATION.get(periodicity, ANNUALIZATION["daily"])
    return sortino_ratio(returns, risk_free_rate) * factor

Example: Comparing Sharpe and Sortino on a Trend-Following Strategy

import numpy as np

# 252 daily returns from a trend-following strategy (positive skew)
daily_returns = np.array([...])  # your actual return array

sharpe = annualized_sharpe(daily_returns.tolist(), periodicity="daily")
sortino = annualized_sortino(daily_returns.tolist(), periodicity="daily")

print(f"Annualized Sharpe:  {sharpe:.3f}")
print(f"Annualized Sortino: {sortino:.3f}")
print(f"Sortino/Sharpe:     {sortino/sharpe:.2f}x  (expect ~1.41 for normal returns)")
# A ratio >> 1.41 confirms positive skew in the return distribution

When Sortino and Sharpe Diverge

The gap between Sortino and Sharpe is a diagnostic tool. For a symmetric (near-normal) return distribution the ratio is ≈ √2. Large deviations signal structural skew:

• Sortino >> √2 × Sharpe (positive skew): The strategy has frequent small losses and occasional large gains — the classic trend-following or long-options profile. Upside runs inflate Sharpe's volatility denominator without representing risk. Sortino correctly ignores them. Use Sortino as the primary metric here.
• Sortino ≈ √2 × Sharpe (near-normal): Consistent with mean-reversion, equity long-only, or diversified multi-strategy. Either metric is reliable; Sharpe is sufficient.
• Sortino < Sharpe (negative excess return): Occurs when the strategy's excess return is negative — the same negative numerator divided by the smaller downside denominator (σ_D ≤ σ) becomes more negative, so Sortino drops below Sharpe. For a losing strategy this is expected, not an error. For a profitable strategy, Sortino < Sharpe would indeed signal a computation bug — check that you used the same annualization factor and denominator convention (n−1) for both.

Sortino vs. Other Risk Metrics

Best practice is to report Sharpe, Sortino, and max drawdown together. Sharpe captures average vol efficiency; Sortino reveals skew direction; max drawdown anchors the actual pain experienced. A strategy with high Sharpe but Sortino << √2 × Sharpe has negative skew and deserves additional scrutiny. AlgoDrill's calculator computes all three simultaneously from your return series.

Failure Modes: When Sortino Misleads

• Too few downside observations: Downside deviation is estimated from the subset of returns below T. If the strategy rarely loses in the sample (e.g., 10 downside observations out of 252), the downside deviation estimate is highly unstable. A Sortino of 8.0 computed on 10 downside observations is statistically indistinguishable from 2.0 — the confidence interval is enormous. Always report the number of downside observations alongside the ratio.
• Regime-conditional downside: A strategy that avoids drawdowns in a bull market may accumulate massive downside deviation in a bear market. A Sortino computed on a single up-trend will not reflect performance across regimes. Use rolling-window Sortino (e.g., 12-month lookback rolling monthly) to track how the ratio evolves across market conditions.
• Target rate sensitivity: Changing T from 0 to the daily risk-free rate (e.g., 0.02% with T-bills at 5%) can shift Sortino significantly. A strategy with many returns just above 0 but below the risk-free rate will show a much worse Sortino under the second convention. Reports without a stated T are not reproducible.
• Look-ahead bias: Identical to the Sharpe failure mode — any use of future data in signal construction dramatically inflates both Sharpe and Sortino. One tick of look-ahead on a high-frequency strategy can produce a Sortino of 20+ from pure noise.
• Short sample: The same Bailey-López de Prado sampling uncertainty that affects Sharpe applies to Sortino. Treat any Sortino computed on fewer than 60 daily observations (with at least 20 downside observations) as a pilot signal, not a firm conclusion.

Sortino and Position Sizing

Like the Sharpe Ratio, Sortino does not directly determine position size — that is the Kelly Criterion's domain. The two metrics complement each other: evaluate whether the strategy has any edge worth sizing (Sharpe ≥ 1.0 or Sortino ≥ 1.4 as a rough gate), then use Kelly to determine the optimal fraction. A high Sortino with a low Sharpe suggests a positively-skewed strategy where Kelly sizing should account for the asymmetric return distribution rather than assuming normality.

The practical sequence: compute Sharpe and Sortino to evaluate strategy quality; use the Sortino-to-Sharpe ratio to diagnose skew direction; apply Kelly with an appropriate distribution assumption; monitor max drawdown to confirm the Kelly fraction is not exceeding psychological limits.