Statistical tools: what each metric tells you and what each hides

Sharpe ratio

Annualized return divided by the volatility of returns. Reward per unit of bumpiness. It is the most quoted number in the business, which is exactly why you should know its blind spots.

It punishes upside the same as downside, so a strategy with occasional big winners scores worse than a steady grinder that made less. It also ignores when the returns happened. One lucky quarter and nine flat ones can carry the same Sharpe as ten even ones. Same number. Different strategy.

Calibration: 0.5 to 1.5 is real-edge territory after costs. Above 2 in retail crypto, be suspicious. Above 3, it is almost always overfitting or market-making. There is no fourth option worth betting on.

Sortino ratio

Sharpe, but only downside volatility counts. Closer to how losing actually feels, since nobody complains about upside surprise. It still hides tails. A strategy with rare disasters posts a lovely Sortino between the disasters. Expect Sortino around 1.2 to 1.5 times the Sharpe. A much bigger gap means the return distribution is strange. Strange distributions deserve a closer look before money.

Profit factor

Gross wins divided by gross losses. Above 1.0 is net profitable. What it hides is composition. One huge winner props up a mediocre strategy’s profit factor for an entire test. Read it next to the trade count, always. A range of 1.3 to 1.8 suggests a real edge. Above 2.5 on hundreds of trades, suspect the fit, not your genius.

Win rate

The most overrated number in trading, and the one beginners chase hardest. It hides almost everything. A 90 percent win rate with 1-to-10 reward-to-risk is a slow funeral: ninety small wins, then ten losses that take it all back. The win rate was real. The equation was negative the whole time.

Win rate means nothing without average win against average loss. Together they are the trader’s equation. Apart they are decoration. Never report one alone. Never chase one alone.

Max drawdown

The worst peak-to-trough drop in the test. This is the most honest number on the panel, because it measures the thing that makes traders quit. It predicts whether you will survive the strategy better than any return figure does.

What it hides is duration. A 20 percent drawdown that recovers in six weeks and one that grinds for six months are different lives with the same headline. Check the underwater time, not just the depth.

Calmar ratio

Annual return divided by max drawdown. Reward per unit of pain. The most underrated metric on the panel, for the same reason drawdown is the most honest. Calibration: 0.5 is decent, 1.0 is good, 2.0 over a long test is excellent and worth double-checking for fit.

Expectancy

Average profit per trade, after costs. The ground truth. It ignores volatility and win rate, which is why it is hard to game. It is the trader’s equation, measured. If expectancy does not clear your per-trade costs with room to spare, nothing else on the panel matters, because the strategy has no business existing.

Pairs that only work together

• Win rate with average win and loss. Apart, misleading. Together, the expected value per trade.
• Sharpe with trade count. A high Sharpe on 12 trades is luck standing up straight. The same Sharpe on 1,200 trades is evidence.
• Sharpe with drawdown. A 1.5 Sharpe with a 40 percent drawdown is a strategy almost nobody can hold. The same Sharpe with 8 percent is usable capital.

What each number bribes the optimizer to do

Hand an optimizer one metric and it finds the degenerate strategy that maximizes it. Sharpe alone buys tight little wins with hidden tails. Win rate alone buys tiny targets and huge stops, the classic blow-up shape. Profit factor alone buys small samples carried by an outlier. Return alone buys leverage. This is the whole argument for multi-objective optimization: metrics in tension keep each other honest. See Pareto frontiers.

Read the panel

The honest report shows Sharpe, Sortino, drawdown, Calmar, profit factor, expectancy, and trade count, with in-sample and out-of-sample versions of each. The reading rule is simple. Spectacular on one metric and mediocre on three is a strategy gaming a number. Solid across the board, with nothing flashy anywhere, is what a deployable strategy looks like. Real edges are boring on every axis. That is how you recognize them.

How Edgecraft handles this

Edgecraft reports the full panel for every backtest, never one headline number, with out-of-sample values beside the in-sample ones and a warning when the gap is wide. Any metric on the panel can serve as an optimization objective, and the advanced analysis view shows what each choice costs the others. See trade quality for the per-fill layer underneath all of these aggregates.