Sharpe vs Sortino vs Calmar: Which Metric to Optimize?

When you are selecting a fitness function for an algo strategy — whether you are running mean-reversion on Solana perpetuals, funding-rate arbitrage on Hyperliquid, or binary markets on Polymarket — the metric you optimize in your parameter sweep becomes the thing your strategy maximizes in production. Choosing between Sharpe, Sortino, and Calmar is not a stylistic preference; it is a load-bearing engineering decision that shapes the entire risk profile of what you ship.

What Each Metric Is Actually Measuring

All three are risk-adjusted return ratios, but they disagree on what "risk" means:

Sharpe Ratio = (mean return − risk-free rate) / standard deviation of returns

Sharpe treats upside volatility the same as downside volatility. A strategy that occasionally prints massive winners gets penalized for it. This is mathematically convenient but economically weird — you are not worried about days where you made 8%.

Sortino Ratio = (mean return − target return) / downside deviation

Sortino replaces the denominator with the standard deviation of only negative returns relative to a target (usually zero). It lets upside volatility run free. In crypto, where return distributions are heavy-tailed and asymmetric, this more closely tracks what you actually care about.

Calmar Ratio = annualized return / maximum drawdown

Calmar divides annualized return by the worst peak-to-trough drawdown over the lookback window (commonly 36 months, though in crypto 12 months is more realistic given regime changes). It says nothing about frequency of losses — only about the single worst run.

Why Optimizing Sharpe Alone Creates Tail-Risk Blind Spots

Here is a concrete example. Consider two hypothetical strategies over 252 trading days:

• Strategy A: 180 positive days averaging +0.15%, 72 negative days averaging −0.18%. Annualized return ~18%, annualized vol ~8.5%, no single drawdown exceeds −6%.
• Strategy B: 200 positive days averaging +0.12%, 50 negative days averaging −0.25%, but with two concentrated loss clusters that produce a −22% drawdown.

Strategy B will often post a higher Sharpe because its overall standard deviation is compressed by all those small wins. But its Calmar is roughly half of Strategy A's, and its Sortino is also worse once you isolate those loss clusters.

If your parameter sweep uses Sharpe as the objective function, an optimizer will find — and prefer — Strategy B. You ship it, it runs fine for six months, and then a regime shift triggers one of those loss clusters at full size. This is not a hypothetical failure mode; it is a documented pattern in systematic fund blowups.

When Sortino Is the Right Default

For most directional and momentum strategies on venues like Hyperliquid, where funding dynamics create persistent positive drift interrupted by sharp reversals, Sortino is the natural default. It rewards you for capturing the drift without penalizing the upside bursts, while still distinguishing strategies that have fat left tails.

Practically, when running a grid search across parameter combinations for a trend-following bot, rank candidates by Sortino and then screen the survivors by maximum drawdown. This two-stage filter catches strategies that look good on Sortino but carry concentrated tail risk — rare but real in thin liquidity windows.

One implementation detail: the target return in the Sortino denominator matters. Using zero means you are measuring deviation below breakeven. Using the risk-free rate or a hurdle makes the denominator smaller on low-volatility strategies, inflating the ratio. Be consistent across your comparison set or your rankings are noise.

When Calmar Is the Right Fitness Function

Calmar becomes the correct primary objective when you are operating under a drawdown mandate — a hard risk limit from a capital allocator, a DAO treasury, or your own position-sizing rules that reference max drawdown as the risk unit. In that regime, optimizing Sortino is solving the wrong problem.

For Polymarket prediction-market strategies, where positions are binary (you lose everything or gain a multiple) and sizing is everything, Calmar forces the optimizer to find parameter sets that survive bad runs rather than just maximizing expected value. A strategy that returns 40% annualized with a 12% max drawdown has a Calmar of 3.3. One that returns 60% with a 35% max drawdown has a Calmar of 1.7 — and it will require a 35% recovery before you are back at high watermark, which in binary-market strategies can take a very long time.

Combining Metrics in Practice

In production, we do not optimize a single metric in isolation. The pattern that works:

• Primary objective: Sortino for return-seeking strategies, Calmar for drawdown-constrained mandates
• Hard constraints: Maximum drawdown cap (e.g., reject any parameter set where max DD > 15%), minimum trade count to ensure statistical significance, minimum holding-period Sharpe floor (e.g., > 0.8) to filter noise
• Tiebreaker: Sharpe for execution robustness — a strategy that scores identically on your primary metric but has lower overall volatility is easier to size and less sensitive to slippage variance

This hierarchy matters because Sharpe is still useful as a robustness proxy, just not as the primary selection criterion.

A Note on Lookback Windows

All three metrics are sensitive to the window you compute them over. In crypto, a 12-month Calmar computed during a bull run is not the same animal as one computed across a full cycle. When backtesting on Solana or Hyperliquid, always compute your metrics across multiple sub-windows — at minimum a trending window, a choppy window, and a high-volatility window — and take the worst-case value as your representative score. Strategies that hold up across all three windows are the ones worth running with real capital.

The algorithmic trading services we build at TierZero are designed with exactly this kind of multi-window, multi-metric fitness evaluation baked into the optimization pipeline, not bolted on afterward.

---

If you are building or auditing a systematic strategy and want a second opinion on your risk metric stack, reach out — this is exactly the kind of problem we work through with clients before a single dollar of live capital is committed.