4. How to evaluate a strategy: drawdown, Sharpe and win rate in context

Theory 5: What algorithmic trading really is

A strategy that is right eight times out of ten can be ruinous. Another that is right three times out of ten can be extraordinary. The hit rate on its own tells you nothing.

With the strategy designed and backtested while avoiding the traps of the previous module, one final question remains: how do you judge whether the result obtained is genuinely good?

We already know the Sharpe ratio from the previous level. In this module we complete the set of essential metrics with maximum drawdown and win rate. And, above all, with the idea that no metric in isolation tells the whole story.

Maximum drawdown: the most visceral metric

Maximum drawdown measures the largest percentage decline a strategy suffers from a peak in capital down to the subsequent trough, before recovering that previous peak.

An example pins it down. An account reaches a value of 100,000 euros, then falls to 70,000 euros, and only afterwards climbs back above 100,000 euros. The maximum drawdown of that period is 30%, regardless of how long the recovery took.

For many managers and investors it is the most visceral metric of all. The reason is that it is the one that best predicts whether a real person will be able to stick with a strategy psychologically to the end.

Think of a strategy with an excellent annualized return but a maximum drawdown of 60%. It may be mathematically attractive on paper and still be practically unviable. Very few real investors can watch more than half their capital disappear without abandoning the strategy at the worst possible moment: right before it recovers.

Win rate: the most misinterpreted metric

Win rate — the percentage of trades that end in a gain — is probably the most misinterpreted metric among beginners.

An 80% win rate sounds spectacular. A 35% win rate sounds mediocre or downright bad. Neither number says anything on its own about whether the strategy is profitable.

What actually matters is expectancy: the win rate combined with the payoff ratio, the relationship between the average size of the gains and the average size of the losses.

The two extremes, in detail

Back to the contradiction we opened with, now with its mechanics:

• A strategy with a 30% win rate can be extraordinarily profitable, if when it wins it wins on average five times more than it loses when it loses. This is the logic behind many trend following strategies, which are right only occasionally but capture large moves on those few occasions.
• A strategy with an 80% win rate can be ruinous, if in the 20% of losing trades the average loss is ten times larger than the average gain on the winning trades.

That second profile is the classic pattern of 'selling options', or of strategies that collect a small premium almost every time, until a rare event wipes out many premiums at once.

Sample size

All of this also requires one final caveat that is forgotten far too often: sample size.

A strategy tested over 15 trades, however good its metrics may look, does not support any statistically solid conclusion. With so few observations, the result may be almost entirely down to chance. No Sharpe or drawdown calculation, however sophisticated, can make up for the lack of data.

As a general and conservative rule, reasonable confidence in a result grows with two things:

• The more independent trades back that result.
• The more distinct market regimes the strategy has lived through during the tested period: bullish, bearish, sideways, high volatility and low volatility.

The more of both, the more reasonable it is to think that the captured pattern is real and not a fluke of the particular period analyzed.

No single figure is enough

No metric in isolation replaces looking at the whole picture, together with the market context in which each result was obtained. Not the Sharpe, not the drawdown, not the win rate, not expectancy.

A solid strategy is not recognized by standing out on a single figure, but by holding a reasonable profile across all of them at once.

And by having demonstrated that consistency over enough data and enough different market conditions to reasonably rule out that all of it was, simply, luck.