Sortino ratio

Two tools can report Sortino ratios that differ by fifty percent from identical data, and neither is wrong. The disagreement is one division, and almost nobody documents which side they took.

The idea

The Sharpe ratio treats all variation as risk, which means an unusually profitable week lowers your score. Traders find that backwards: nobody is harmed by upside.

Sortino keeps the structure and replaces the denominator. Instead of the deviation of all days, it uses the deviation of the losing days only:

Sortino = (mean daily result ÷ downside deviation) × √252

Everything else matches our Sharpe implementation — daily sums of net P&L, currency rather than returns, no risk-free rate subtracted, and the days you did not trade absent from the series entirely. Those three assumptions are explained on the Sharpe page and they apply here unchanged.

The disagreement: what you divide by

Downside deviation is a root-mean-square of the negative days. The dispute is the count underneath:

Convention A — divide by all periods.

downside deviation = √( Σ (negative days)² ÷ N total days )

Convention B — divide by losing periods only.

downside deviation = √( Σ (negative days)² ÷ N losing days )

Because the losing days are a subset, convention B always divides by the smaller number. That makes the downside deviation larger, and therefore the Sortino ratio smaller, whenever there was at least one winning day.

The gap is not cosmetic. On a series where a third of days lose, convention B's denominator is roughly 1.7 times convention A's, and the reported ratio is about 40% lower from exactly the same trades.

Ours uses convention B — the average is taken over losing days. It is the more conservative reading, and it answers a question a trader can actually feel: on a bad day, how bad is bad, relative to what a typical day earns. Convention A instead blends that with how often bad days happen, which is information the ratio does not need because the mean in the numerator already carries it.

Neither is a mistake. What is a mistake is comparing a figure from one convention against a figure from the other, which happens constantly because most tools state neither.

The target is zero

Both conventions need a threshold below which a day counts as downside. The general definition allows any minimum acceptable return; ours uses zero — a losing day is a day that lost money.

Setting the target above zero (say, a required daily gain) turns break-even days into downside and lowers the ratio. That is a legitimate variant and a different measurement, and it is worth knowing that a ratio quoted without a stated target is ambiguous.

The edge case: no losing days

If the account has no losing days, the denominator is zero and the ratio is undefined. The three ways of handling that are the same three that profit factor faces, and the choice is identical here: ours returns no value rather than infinity or a large placeholder.

This comes up more than expected on short samples. A first week of four profitable days produces no Sortino ratio at all, and that is the honest output — four days is not evidence of anything, and a number would imply otherwise.

Reading it against Sharpe

The two together say something neither says alone:

  • Sortino much higher than Sharpe — your variation is mostly on the winning side. Large winners, contained losers. Usually a good sign, occasionally a sign that one outsized winner is carrying the record.
  • Sortino close to Sharpe — losses and gains vary about equally.
  • Sortino below Sharpe — the variation is concentrated in the losses. Under convention B this needs care, since the conservative denominator pushes the figure down on its own; check whether the losing days are genuinely more scattered than the winning ones before reading it as a finding.

Neither ratio sees order. A run of losses arriving consecutively scores the same as the identical losses spread across a year, and only maximum drawdown and the losing streak view read the sequence.

More in Trading terms, defined by how they are computed

  • Net P&LThe result of a trade after commission and swap, and why the sign convention in broker exports makes double-counting so easy.
  • Profit factorGross profit divided by gross loss, the edge case that breaks it, and why a high profit factor on few trades means almost nothing.
  • ExpectancyThe expected value of one trade, the break-even win rate it implies, and why the figure needs an error bar to mean anything.
  • R-multipleExpressing results as multiples of the amount risked, why it survives account growth, and the case where R stops being comparable.
  • Win rateWhat share of trades finished positive, how break-even trades are counted, and why the figure is uninterpretable without the win-to-loss ratio.
  • Maximum drawdownThe largest peak-to-trough fall in your account, and the measurement choice that decides whether a prop account survives.
  • Trailing drawdownA loss limit that rises with your account and usually never falls back, plus the two sentences in a rulebook that decide when it can end your account.
  • Consistency ruleA cap on how much of your profit may come from a single day or trade, why it exists, and how it turns a winning account into an unpayable one.
  • Payoff ratioThe size relationship between a typical winner and a typical loser, how break-even trades drop out of both averages, and why the realised ratio rarely matches the planned one.
  • SQN (System Quality Number)Mean R divided by the standard deviation of R, multiplied by the square root of the sample, and the two reasons the score is not comparable between accounts.
  • Sharpe ratioMean daily result divided by its standard deviation and annualised, plus the three assumptions that make a trading journal's Sharpe different from a fund's.
  • Average holding timeHow long a typical trade stays open, why open positions are excluded from the average, and the reason a disposition-effect account reports a short one.