Tick, Volume And Dollar Bars

A standard 1-minute bar samples the market once every 60 seconds of clock time, regardless of whether 5 trades happened in that minute or 5,000. Right after a big news release, a huge amount of genuinely new information gets crammed into a handful of clock minutes, and a fixed-time bar averages right over it; during a quiet lunchtime lull, a 1-minute bar is sampling almost nothing, adding mostly noise. Clock time is not the clock that markets actually run on — activity is the clock that matters.

Tick bars close after a fixed number of trades have occurred (e.g., every 500 trades), regardless of how much time that took. Volume bars close after a fixed number of shares have traded (e.g., every 100,000 shares). Dollar bars close after a fixed dollar value has traded (e.g., every $5,000,000 notional) — the preferred choice for most research because it naturally adjusts for a stock's price level and share count changing over time (a 2-for-1 split doubles share volume but leaves dollar volume unaffected, so dollar bars need no adjustment when a corporate action changes the share count).

Worked example. A liquid stock trades roughly 2,000,000 shares in a typical day at an average price near $50, so daily dollar volume is about $100,000,000. Choosing a $1,000,000 dollar-bar threshold produces roughly 100 bars on a typical day — but not evenly spaced in time. In the first 30 minutes after the open, where a third of the day's volume often trades, dollar bars might close every 20–40 seconds; during a quiet mid-afternoon stretch, the same $1,000,000 threshold might take 10–15 minutes to fill. The bar count adapts to activity automatically, with no need to separately model when the market is busy.

Second example, why this matters for correlation. Two 1-minute return series measured in clock time will show spurious low correlation around news events, because the news hits at slightly different clock-minute boundaries relative to when each stock actually reacted. Rebuilding the same data as $2,000,000 dollar bars for both names causes each series to sample more densely exactly during the shared burst of activity, more closely aligning the bars in "activity time" even though they remain offset in clock time — improving the measured correlation's stability without changing anything about the underlying prices.

What this means in practice

Machine-learning pipelines for trading signals generally prefer dollar or volume bars over clock-time bars because the resulting return series is closer to independently and identically distributed — a core assumption behind most statistical tests and ML training procedures — since each bar represents roughly the same amount of "information absorbed" rather than the same amount of elapsed time. The tradeoff is operational complexity: bar boundaries change in real time and need to be computed on the fly rather than simply ticking on a clock.

See Information-Driven Bars for a further refinement that samples based on order-flow imbalance rather than raw activity, and The Intraday Volume Profile for how volume itself varies predictably across the trading day even before bars are built from it.