Anatomy of a real trading edge

A real trading edge has three properties: it is statistically meaningful, it survives realistic transaction costs, and it holds up out-of-sample on data the researcher has never touched. Most published 'edges' fail at least one of these tests, and a surprising number fail all three when looked at honestly.

Statistical significance is the easy one. Cost survival is harder and the place most retail research falls apart — a flat per-trade cost assumption is almost always too generous, and the difference between a flat-cost model and a realistic one is often the difference between profitable and not. Out-of-sample robustness is the hardest, because by definition you only get one out-of-sample window, and once you have looked at it, it stops being out-of-sample.

We are deeply suspicious of strategies with high headline Sharpe ratios on short histories. A two-year Sharpe-3 backtest is almost always overfit to the regime that produced it. The strategy looks like it works because the regime cooperated, not because the strategy is good.

We weight long, ugly, drawdown-heavy backtests on multi-decade data far more than clean recent ones. A strategy that has earned a Sharpe of 0.8 across decades while surviving 2008, 2020, and several smaller crises has communicated something durable about its edge. A strategy that has earned a Sharpe of 3 since 2022 has communicated something about the post-2022 regime — and not much else.

Walk-forward analysis is the standard tool for assessing out-of-sample behaviour, and it is widely misused. The version that matters is genuinely sequential: train on a window, test on the next disjoint window, advance, repeat. No peeking. No reusing the same out-of-sample window for hyperparameter tuning and final evaluation.

Done properly, walk-forward kills many strategies that look excellent on a single in-sample fit. That is the entire purpose of the exercise. A strategy that survives genuine walk-forward across multiple regimes is exponentially more likely to survive live deployment than one that has only survived in-sample.

When a strategy survives the research gauntlet, it does not get deployed at theoretical optimum. It gets deployed at a small fraction — typically a quarter or less — of its modelled sizing. The live equity curve is then compared to the backtest equity curve over a defined live period, with explicit tolerance bands.

If live performance tracks backtest within tolerance, sizing is scaled up incrementally. If it does not, the strategy is paused, debugged, and either redeployed at the small size or killed. The boring truth is that most strategies that look great on paper never make it past this stage, and the ones that do are the ones we end up trading at meaningful size.

The edges that last in our experience share a profile. They are usually small in headline terms — a few basis points per trade, an information ratio under 1.5, a hit rate barely above 50%. They are usually easy to explain in one sentence. And they are almost always unglamorous to implement, with most of the engineering cost going into execution, cost control, and risk management rather than into the signal itself.

This shape is unsatisfying to discuss in marketing material, which is one of the reasons it is rarely discussed in public. It is, however, the shape of the edges we actually trust, and we have not seen a counterexample worth mentioning.