Microstructure noise: why second-by-second charts are mostly nonsense
Posted: Sat Oct 31, 2026 8:19 pm
At the finest time scales, price movement is dominated by effects unrelated to value. Quote updates, rounding, spread changes and algorithmic adjustments create jitter.
A chart built on second-by-second data looks active, yet most of the movement is noise that reverses almost immediately. Indicators on that data respond to the noise.
For scalpers, that suggests caution about too-fast timeframes. Signals generated from one-second candles have a low signal-to-noise ratio, and the spread alone is bigger than many of the moves.
Filters help. Longer smoothing, such as using the median of several seconds, or sampling at fixed intervals, can reveal the underlying movement.
Another approach is to check how a signal's predictive power changes with the timeframe. If it only works at one specific scale and vanishes at neighbouring ones, it may be an artefact.
The underlying lesson is that faster isn't automatically better. There's a point where noise outweighs information.
One useful experiment is to plot the same signal's profitability against sampling interval, from one second up to five minutes. A real edge usually stays positive across several neighbouring intervals, while a noise artefact appears at one value and nowhere else.
What is the shortest timeframe you trust for signals?
A chart built on second-by-second data looks active, yet most of the movement is noise that reverses almost immediately. Indicators on that data respond to the noise.
For scalpers, that suggests caution about too-fast timeframes. Signals generated from one-second candles have a low signal-to-noise ratio, and the spread alone is bigger than many of the moves.
Filters help. Longer smoothing, such as using the median of several seconds, or sampling at fixed intervals, can reveal the underlying movement.
Another approach is to check how a signal's predictive power changes with the timeframe. If it only works at one specific scale and vanishes at neighbouring ones, it may be an artefact.
The underlying lesson is that faster isn't automatically better. There's a point where noise outweighs information.
One useful experiment is to plot the same signal's profitability against sampling interval, from one second up to five minutes. A real edge usually stays positive across several neighbouring intervals, while a noise artefact appears at one value and nowhere else.
What is the shortest timeframe you trust for signals?