Select the Region, Not the Peak Cell

When I tune a strategy in Trident Forge, the number I trust least is the best one on the grid.

Here’s the discipline, and it predates any chart I’ll cite: I don’t pick the cell with the highest backtested Sharpe. I pick the region that holds up around it, and let the cell fall out of that. A peak that stands alone is a peak I can’t stand on with live money — it’s the parameter set that fit the noise in my sample, and noise doesn’t repeat. So I score each cell by its neighborhood, not its coordinate: mean(3×3) − std(3×3). Reward the plateau, penalize the cliff. The region wins; the cell just comes along for the ride.

That’s the habit. What made it click as more than a hunch was an illustration @rational_edge put up — “The Cell.” [1] He ran a 2,760-cell SMA-crossover grid on the S&P 500 and took the “best” cell — Sharpe 0.91 — apart. It split three ways: 0.61 belonged to the strategy family (any cell earns roughly that), 0.27 to the 3×3 neighborhood around the peak, and only 0.04 to the exact tuned cell. The thing that feels like the win — the fine-tuned coordinate — was four hundredths of a Sharpe. The rest was the family and the region.

That’s the part worth keeping. You’re not hunting a magic coordinate. You’re checking whether a neighborhood exists that pays out no matter which cell inside it you happen to land on.

And there’s a tell worth watching for — and it’s right there in @rational_edge’s own numbers. The shorter the backtest window, the more of that Sharpe loads into the exact cell. A short sample doesn’t hold enough real signal to reward a whole region, so the optimizer finds a coordinate that fit the wiggles. If your best number only lives at one setting, and it gets lonelier as you shorten the window, you didn’t find an edge. You found the sample.

None of this is a discovery of mine — robust optimization is old, and the decomposition is @rational_edge’s, not mine. What’s mine is the habit: I’d rather run from the dull middle of a broad plateau than the bright top of a narrow spike, every time.

The backtest was never the edge. The region is.

References

[1] @rational_edge, “The Cell,” Instagram.