Most backtest-to-live performance degradation isn't caused by overfitting at all—it's caused by the implicit assumption

Your strategy didn't overfit. Your execution model lied to you.

The quant community has developed a near-reflexive diagnosis for backtest-to-live performance degradation: overfitting. Too many parameters. Too little out-of-sample discipline. Insufficient regularization. But here is what I find difficult to reconcile with that narrative. Even rigorously validated strategies with minimal parameter counts, strategies that pass every combinatorial purged cross-validation test you can throw at them, still experience 30 to 60 percent Sharpe ratio degradation when they go live. That gap is too large and too consistent to be explained by overfitting alone. Something else is compounding against you. And almost nobody is measuring it.

The dominant framework in quantitative finance treats backtest-to-live degradation as fundamentally a statistical problem. Marcos López de Prado's *Advances in Financial Machine Learning* has become something close to canonical, and for good reason. His work on the Deflated Sharpe Ratio, combinatorial purged cross-validation, and the dangers of backtest overfitting gave the industry rigorous tools to address a real problem. Platforms like QuantConnect and Zipline have built walk-forward optimization into their core workflows. Institutional risk systems audit parameter stability and out-of-sample decay as standard practice. The entire diagnostic apparatus of modern quant finance points at one villain: statistical overfitting. This persists because overfitting is intellectually tractable. You can test for it. You can publish papers about it. You can build products around it. But tractability is not the same as completeness.

The common view is dangerously incomplete because it treats the execution layer as a neutral pass-through. A transparent pane of glass between signal and PnL. Your model says buy at price P, the backtest records a fill at price P, and the walk-forward engine evaluates alpha at price P. In production, none of this holds. Consider what actually happens when your strategy places a resting limit order at the best bid on ES futures. Your backtest fills that order the moment price touches your level. In the live order book, you are not alone at that level. You are behind potentially thousands of contracts placed by faster participants with collocated infrastructure and queue priority algorithms purpose-built to be ahead of you. Empirical data from the CME reveals that the probability of a limit order at the best bid actually filling when price merely touches that level can be as low as 15 to 25 percent. Your backtest assumes 100 percent. That is not a rounding error. That is a 3x to 6x overstatement of fill rate, applied to every single trade in your simulation.

Now extend this across the full surface of execution uncertainty. Partial fills mean your position sizing is stochastic, not deterministic. Variable latency means your execution timestamps are drawn from a jitter distribution, not a fixed point. The market impact of your own orders, even at modest participation rates, shifts the price against you in ways that are correlated with your signal's conviction. And critically, every one of these effects is invisible to any in-sample or out-of-sample statistical test. Walk-forward analysis cannot detect a problem that exists in the execution layer because walk-forward analysis does not model the execution layer. It models alpha stability. You are testing the wrong surface.

This problem compounds because the execution surface itself is non-stationary. The Almgren-Chriss optimal execution framework demonstrates that market impact costs scale nonlinearly with both volatility and participation rate. During the exact volatility clusters that generate your strongest signals, execution costs spike, fill rates collapse, and adverse selection intensifies. The variance drag is not a fixed cost you can subtract. It is a shifting tax that increases precisely when your strategy needs its edge most. I have seen this firsthand. A mid-frequency stat-arb strategy backtested at a 2.1 Sharpe degraded to 0.8 in live trading. Forensic decomposition attributed over 70 percent of the gap not to alpha decay or parameter overfitting but to fill-rate assumptions and latency-driven adverse selection. The alpha was real. The execution model was fiction.

Sophisticated practitioners address this not by adding a fixed slippage assumption, which is just another form of false precision, but by building stochastic execution simulators. The architecture has four components. First, a queue position model calibrated on historical Level 3 or inferred queue data. Second, a latency injection layer that randomizes execution timestamps within empirically measured jitter distributions. Third, a fill probability model conditioned on trade-through frequency and order book depth. Fourth, a post-trade analytics module that performs Brinson-style attribution decomposing PnL divergence into signal alpha, execution capture, and residual microstructure noise. Firms like Two Sigma and Citadel Securities have publicly stated they invest as much engineering effort in execution simulation as in alpha research. That ratio tells you where the real edge leakage lives.

If execution non-determinism is the dominant driver of backtest-to-live degradation, then our collective obsession with ever more sophisticated overfitting diagnostics may itself be a form of overfitting. We are optimizing our tools for a problem that explains the minority of realized performance drag while ignoring the one that compounds silently beneath it. So here is my challenge to every systematic practitioner reading this. Quantify, with forensic certainty, what percentage of your backtest-to-live PnL gap is attributable to alpha overestimation versus execution model error. I predict most of you will find you have never performed that decomposition. An unexamined assumption embedded in every trade you simulate is the most dangerous kind of overfitting there is.