How to Compare Trading Models — Return Is the Least Interesting Column

How to compare trading models comes down to one rule: never rank them on a single number. Two models can post the same return while being entirely different in quality, because one earned it steadily and the other earned it in one lucky stretch. A fair comparison puts return, risk and consistency side by side.

Compare like with like first

Before looking at any metric, check that the models are being judged under the same conditions. In darwintIQ every model is evaluated on the same rolling four-hour window, so the figures on the dashboard are directly comparable. That matters more than it sounds. A model measured over a quiet Asian session and another measured through a news spike are not facing the same test, and any comparison between them says more about the calendar than about the models.

Also compare models on the same symbol. A strategy that suits the characteristics of one instrument may behave very differently on another, so a ranking that mixes symbols mostly measures the instruments.

Read return against the cost of earning it

Return is the headline, but it is the least informative column on its own. Pair it with the figures that describe how the return was earned. Drawdown shows the worst stretch the model endured on the way. The Sharpe ratio and Sortino ratio relate return to volatility and to downside volatility respectively. The Calmar ratio relates return to the deepest drawdown.

A model with a modest return and a shallow drawdown will often beat a flashier rival once risk is included. The question to ask of every candidate is not how much it made, but how uncomfortable it was to hold while it made it.

Check consistency, not just the average

An average hides its own spread. Profit Factor and Expected Value tell you what a typical trade is worth, but they say nothing about whether that value is stable. This is where the consistency measures earn their place: the Robustness Score, the Stability Score and Return Stability all ask whether the result is repeatable rather than whether it is large.

When two models are close on return, the one with the better consistency scores is usually the safer choice. The reasoning is covered in more detail in the article on the Robustness Score.

Use the Trader Detail view to see the structure

The numbers are only half the picture. Open each candidate in the Trader Detail view and look at its construction: the entry logic, the position manager and the regime filter. Two models with similar metrics can rest on very different foundations. A trend-following entry with an ATR position manager and a trend regime filter is built for a different market from a range-bounce model with a support and resistance filter.

Ask whether the design suits the current market regime. A model that ranks highly while the regime matches its design may fall away quickly when conditions change, which is why the regime context on the dashboard belongs in the comparison. The TrendMatrix helps here, since it shows direction and strength across timeframes at a glance.

Watch the trade count

A small number of trades makes every metric unreliable. A model showing an excellent Sharpe ratio on a handful of trades has told you very little, because a few lucky outcomes can produce almost any statistic. Prefer the model whose results rest on enough trades to mean something, and treat a spectacular figure on thin evidence with suspicion.

A short comparison checklist

When time is short, run through five questions in order. Are the evaluation conditions identical? Is the drawdown acceptable for the return? Do the consistency scores support the headline figures? Does the design suit the current regime? Is the trade count high enough to trust? A model that answers yes to all five deserves attention. One that fails two or more rarely repays it, whatever its return.

Final thoughts

Comparing trading models well is mostly a matter of discipline. Use the same conditions, read return against risk, favour consistency over the peak, inspect the design and respect the sample size. None of these steps requires a complicated tool; they require resisting the pull of the biggest number on the screen. The model worth following is rarely the one at the top of a single column, but the one that holds up across all of them.