Microstructure noise can swamp one-second signals

[Amanhasnoname]

At very short time scales, the price that gets recorded is the true value plus noise from bid-ask bounce, quote flickering, and rounding. Over seconds, noise can be larger than the genuine movement.

Anyone building a signal from one-second bars gets a lot of false patterns. Indicators respond to every flicker, and the resulting trades are mostly noise capture, which costs a spread each time. Backtests look decent because the noise is symmetric and the test ignores slippage, but live trading collapses.

Mitigation options exist. Sampling at a slightly longer interval reduces noise. Using mid prices or smoothing quotes removes some bounce. Requiring a minimum move before reacting avoids chasing flicker.

Equally important is a realistic cost model. At one-second frequency, every trade pays the full spread, and any signal weaker than that cost is worthless.

The honest question is whether the target time frame even holds a tradeable signal, or whether the apparent patterns are all measurement artifacts.

What is the shortest time frame that has produced a signal surviving realistic costs?