Why optimizing for return alone overfits

A clear conceptual case: optimizing on raw return alone increases backtest overfitting because it rewards leverage, variance, and lucky in‑sample outcomes, while risk‑adjusted targets such as the deflated Sharpe ratio better account for volatility and multiple testing. However, there is not enough empirical detail to substantiate stronger quantitative claims about exact out‑of‑sample degradation, leverage differences, or comparative case‑study results.

Optimizing a strategy on raw return sounds intuitive, but validation literature argues that it systematically points search toward the wrong kinds of rules. In Sonar’s strategy-validation framework, backtest overfitting happens when a research process selects parameterizations or rules that fit historical noise rather than persistent structure. A key reason is that raw return is highly sensitive to leverage, concentration, and a small number of favorable path outcomes, so a search procedure that rewards return alone tends to discover those exposures before it discovers robust signal quality.

The strategy-validation material emphasizes that a backtest is not just an estimate of opportunity; it is also a record of how much variance, path dependence, and model-selection freedom were used to produce that estimate. When researchers try many variants, the best in-sample result is often the one that benefited most from luck. In that setting, optimizing for raw return amplifies selection bias because return does not penalize the instability introduced by leverage or the fragility introduced by a small number of outsized trades. The result is a strategy candidate that can look compelling in-sample while degrading materially out of sample.

The backtest-overfitting audit tool is built around this exact concern: evaluating whether the model-selection process has likely chosen a rule that will fail to generalize. The tool’s framing is that search itself creates false discoveries, and the more aggressively an objective rewards extreme in-sample outcomes, the more likely the selected specification is to be a lucky one. Within that logic, a return-maximizing objective is especially exposed because it can rank a highly levered, high-variance rule above a more stable one even when the latter has better evidence of repeatability.

This is the motivation for using risk-adjusted targets. The glossary entry on the deflated Sharpe ratio describes it as a Sharpe-based statistic adjusted for non-normal returns and multiple testing, intended to measure whether an observed Sharpe is truly exceptional once selection effects are considered. That adjustment matters for strategy search. A target that incorporates risk and the number of trials does not simply ask, “Which specification made the most money in sample?” It asks a stricter question: “Which specification delivered risk-adjusted performance strong enough to remain meaningful after accounting for skew, kurtosis, and the fact that many alternatives were tried?”

That distinction changes what the optimizer prefers. A raw-return objective can be improved mechanically by taking more exposure or by loading onto episodic market conditions. A risk-adjusted target such as a deflated Sharpe ratio is less willing to reward those choices unless the excess performance survives volatility and multiple‑testing penalties. In practice, that means the search is pushed away from fragile parameter sets that win by scale or timing luck and toward rules whose edge is more proportionate to the risk and search complexity used to obtain it.

The defensible conclusion is that if a research process optimizes only for return, it increases exposure to backtest overfitting because raw return is easily inflated by leverage, favorable noise, and model‑selection luck. A validation target that prices in risk and corrects for multiple testing, such as the deflated Sharpe ratio, is better matched to the goal of identifying strategy rules that generalize across unseen data and changing regimes.

Drafted with AI assistance from cited sources. Reviewed and approved by Sonar Sciences Quant & Research Team.