Sharpe Ratio: Risk-Adjusted Return Benchmark
The Sharpe ratio measures excess return per unit of volatility, calculated as (Strategy Return − Risk-Free Rate) / Standard Deviation of Returns. A Sharpe of 1.0 indicates acceptable risk-adjusted performance; above 2.0 is excellent for daily-return strategies, though values above 3.0 in backtests often signal overfitting or data errors. Critical nuance: Sharpe penalizes upside volatility identically to downside volatility, which misrepresents asymmetric strategies like trend-following or options selling. Always compute Sharpe using the same return frequency as your rebalancing schedule—annualizing daily Sharpe by √252 inflates values when returns are autocorrelated. Compare your Sharpe against the strategy’s benchmark Sharpe over identical periods, not against zero. Negative Sharpe ratios are not interpretable in magnitude because the denominator remains positive while the numerator flips sign; a Sharpe of −0.5 is not “half as bad” as −1.0. Report Sharpe with its standard error (approximately √((1 + Sharpe²/2)/N)) to judge statistical significance. A Sharpe of 1.5 over 36 monthly observations has a standard error near 0.42, meaning the true Sharpe could plausibly be 0.7 or 2.3.
Sortino Ratio: Downside-Only Risk Adjustment
The Sortino ratio replaces total standard deviation with downside deviation, calculated only from returns below a minimum acceptable return (MAR), typically zero or the risk-free rate. Formula: (Strategy Return − MAR) / Downside Deviation. Sortino exceeds Sharpe for strategies with positively skewed return distributions—trend-following, long volatility, and catastrophe bonds—because large positive outliers no longer inflate the denominator. A strategy with Sharpe 0.8 and Sortino 1.6 has materially better downside protection than one with Sharpe 1.0 and Sortino 1.1. Interpret Sortino only alongside the MAR definition; changing MAR from 0% to 2% annualized can shift Sortino by 30–50% for low-return strategies. Downside deviation using only negative returns (target 0%) versus returns below the mean produces different values—state your convention explicitly. Sortino is unstable with fewer than 24 observations because downside deviation requires sufficient negative-return samples; with 12 monthly returns and only 3 negative months, Sortino is essentially noise.
Maximum Drawdown: Peak-to-Trough Capital Erosion
Maximum drawdown (MDD) is the largest percentage decline from a running peak to a subsequent trough, expressed as a positive number: MDD = (Trough Value − Peak Value) / Peak Value. MDD answers the question every allocator asks: “How much capital could I have lost at the worst moment?” A 25% MDD requires a 33% gain to recover; a 50% MDD requires 100%. Interpret MDD relative to strategy leverage and volatility targeting—a 20% MDD on 2x leveraged exposure equals 10% unlevered. MDD is path-dependent and sample-specific: a 10-year backtest captures the 2008 and 2020 crises, but a 3-year backtest may show artificially low MDD. Always report MDD alongside the drawdown duration (time from peak to recovery) and the Calmar ratio (annualized return / MDD). Calmar above 0.5 is respectable; above 1.0 is strong. Beware backtests reporting MDD below 10% for equity-like strategies—this usually indicates survivorship bias, look-ahead bias, or a sample period lacking stress events. Stress-test MDD by applying the strategy’s beta to historical crisis returns; if your backtest MDD is 8% but your strategy’s implied beta to the S&P 500 is 0.9, the 2008 S&P drawdown of 55% implies a 50% expected MDD in a repeat scenario.
Volatility and Annualized Standard Deviation
Annualized volatility equals the standard deviation of periodic returns multiplied by √periods per year (√252 for daily, √12 for monthly). Volatility contextualizes every other metric: a 15% annualized return with 10% volatility (Sharpe 1.5 assuming 0% risk-free) differs fundamentally from 15% return with 30% volatility (Sharpe 0.5). Examine volatility clustering—GARCH effects mean volatility is autocorrelated, so a single annualized figure masks regime shifts. Compute rolling 63-day volatility and check whether backtest volatility matches realized volatility in live trading; a persistent gap signals execution slippage or model drift. Compare strategy volatility to its benchmark: a market-neutral strategy with 8% volatility is high-risk relative to peers at 4%, even though 8% looks modest versus equities. Report both full-period and downside volatility; a ratio of downside to total volatility above 1.2 indicates meaningful negative skew.
Beta and Alpha: Separating Skill from Exposure
Beta measures sensitivity to benchmark returns via regression: Strategy Return = Alpha + Beta × Benchmark Return + Error. Beta of 1.2 means a 10% benchmark move implies a 12% strategy move. Alpha is the intercept—return unexplained by benchmark exposure. A strategy with 12% annual return, beta 1.0, and benchmark return of 11% has 1% alpha; the same 12% return with beta 0.3 and benchmark at 11% has 8.7% alpha, revealing genuine skill. Interpret alpha only with its t-statistic; alpha of 5% annually with a t-stat of 0.8 over 5 years is statistically indistinguishable from zero. Use Newey-West standard errors when returns are autocorrelated, and regress against multiple factors (Fama-French three-factor or five-factor) to avoid misattributing factor premia—value, size, momentum, profitability, and investment—to alpha. A long-only small-cap value strategy will show large “alpha” versus the S&P 500 that vanishes against a small-cap value benchmark. Report R²: alpha from a regression with R² of 0.85 is mostly residual noise; alpha from R² of 0.15 reflects genuine idiosyncratic return.
Win Rate, Profit Factor, and Payoff Ratio
Win rate equals winning trades divided by total trades. A 40% win rate with a 2.5:1 average win/loss ratio (profit factor 1.67) outperforms a 65% win rate with a 0.5:1 ratio (profit factor 0.93). Profit factor = gross profits / gross losses; above 1.5 is solid, above 2.0 is excellent, and below 1.0 is unprofitable. Payoff ratio = average win / average loss. Interpret these three jointly: win rate × payoff ratio determines expectancy, and expectancy × trade frequency determines total return. A high win rate with tiny wins and occasional catastrophic losses (negative skew) produces smooth backtests that collapse in live trading—check the win-rate distribution across market regimes. Compute win rate excluding the top 5% of trades; if win rate drops from 55% to 38%, the strategy depends on rare outliers and has fragile edge. Require at least 100 trades for stable win-rate estimates; 30 trades give a win-rate confidence interval of roughly ±18 percentage points.
Turnover, Capacity, and Transaction Cost Sensitivity
Turnover measures portfolio replacement per period: annual turnover of 300% means the average holding period is four months. High-turnover strategies look attractive gross but erode under realistic costs. Model costs explicitly: commission (1–5 basis points), bid-ask spread (2–50 bps depending on liquidity), market impact (scales with position size relative to ADV), and slippage. Recompute Sharpe after subtracting 2× turnover × round-trip cost; a gross Sharpe of 1.8 can fall to 0.9 at 400% turnover with 15 bps round-trip costs. Capacity estimates the AUM at which market impact consumes the strategy’s edge—strategies trading illiquid small caps may cap at $50 million while large-cap strategies sustain billions. Test cost sensitivity by doubling assumed costs; strategies whose profitability vanishes under 2× costs are fragile. Report breakeven cost—the per-trade cost at which net return equals zero—as a capacity diagnostic.
Statistical Significance and the Deflated Sharpe Ratio
Backtest metrics without significance testing invite false discovery. With 50 strategy variants tested, the best Sharpe of 1.4 is expected even if all have zero true edge (multiple testing). The Deflated Sharpe Ratio (Bailey & López de Prado) adjusts for the number of trials, non-normality, and sample length: DSR = Φ(((Sharpe − SR₀) × √(N−1)) / √(1 − γ₃×Sharpe + ((γ₄−1)/4)×Sharpe²)), where SR₀ is the expected maximum Sharpe under the null across K trials. Require a t-statistic above 2.0 on excess returns, and apply the Bonferroni or Holm correction when testing many variants. Bootstrap the return series (1,000 resamples) to generate confidence intervals for Sharpe, MDD, and profit factor—report the 5th and 95th percentiles, not just point estimates. A backtest Sharpe of 1.3 with a bootstrap 5th-percentile Sharpe of 0.2 is not robust enough for capital allocation. Check for regime dependence by computing Sharpe in rolling 24-month windows; a strategy with full-period Sharpe of 1.2 but only 3 of 10 windows above 1.0 shows unstable edge.
Benchmark Selection and the Information Ratio
The information ratio (IR) = active return / tracking error, where active return is strategy return minus benchmark return and tracking error is the standard deviation of that active return. IR above 0.5 is good; above 0.75 is very good; above 1.0 is exceptional and rare. Benchmark choice determines IR entirely: a long-only equity strategy measured against the S&P 500 may show IR of 0.4, while the same strategy against a custom factor blend shows IR of 0.1. Use the benchmark the strategy’s mandate implies—not a convenient index. Report correlation to the benchmark; correlation of 0.95 with IR of 0.6 means the strategy is mostly beta with a thin active overlay. Examine active return attribution: how much comes from sector allocation, security selection, and factor tilts? A strategy claiming selection skill but showing 90% of active return from a single sector bet has concentration risk, not skill. Compute IR over rolling 36-month windows; IR stability matters more than full-period IR, which can be dominated by one strong year.
Metric Interdependence and Composite Scoring
Metrics overlap and can contradict: high Sharpe with severe MDD (negative skew), strong alpha with high turnover (cost-fragile), high win rate with low profit factor (small wins, large losses). Build a composite scorecard weighting metrics by strategy type—trend-following weights Sortino and Calmar heavily; market-neutral weights Sharpe, IR, and drawdown duration; options strategies weight tail-risk metrics like skewness and maximum drawdown. Require minimum thresholds across all dimensions: Sharpe > 0.7, Calmar > 0.4, DSR > 0 with 95% confidence, profit factor > 1.3, and breakeven cost > 2× assumed cost. Reject strategies failing any threshold regardless of strength elsewhere—a 2.0 Sharpe with a 60% MDD is uninvestable at typical risk limits. Track metric stability across sub-periods and parameter perturbations; a strategy whose Sharpe swings from 0.5 to 2.0 under ±10% parameter changes is overfit. Finally, compare backtest metrics to live or paper-trading metrics over identical horizons; degradation of more than 30% in Sharpe or 50% in profit factor signals implementation shortfall requiring diagnosis before capital deployment.
