Volume bars, dollar bars and the problem with the clock

August 14, 2026

A five-minute bar at the open and a five-minute bar at lunch are the same duration and nothing else. Sampling by clock time oversamples quiet periods and undersamples the ones that matter, and it breaks the statistical assumptions most indicators quietly rely on.

Almost every chart you have ever looked at samples the market by clock time. One bar per minute, per hour, per day. It is so standard that the choice is rarely recognised as a choice.

It is a choice, and it is a poor one for most analytical purposes.

The problem

Markets do not produce information at a constant rate.

The first five minutes after the open can carry more trading than the next ninety. A quiet mid-afternoon stretch might produce a few hundred contracts. A five-minute bar spanning each of those is the same duration and nothing else: one is a dense record of intense activity, the other is mostly the absence of anything.

Time-based sampling therefore does two harmful things at once. It oversamples quiet periods, generating bars with almost no information, and it undersamples active periods, compressing the most eventful trading into the same single observation as the dullest.

The consequences are statistical, and they are the ones that bite.

Returns from time bars are strongly heteroscedastic. Variance changes dramatically across the session in a predictable, structural pattern. Many common techniques assume something closer to constant variance.

They exhibit serial correlation driven by the activity cycle rather than by anything about the market.

They are further from normally distributed than the alternatives, and a great deal of standard statistical machinery is calibrated on distributions closer to normal than intraday time bars actually are.

Volume bars

The alternative is to sample by activity. A volume bar closes when a fixed quantity has traded, whether that takes three seconds or two hours.

The result is that every bar contains a comparable amount of trading. Busy periods generate many bars, quiet periods generate few, and the sampling density adapts to what the market is doing rather than to the clock.

This is the same reasoning behind VPIN's volume clock, and it comes from the same body of work. The empirical result reported in that literature is that volume bars produce return series with better statistical properties: closer to normal, less heteroscedastic, less serially correlated.

Dollar bars

Dollar bars extend the idea by sampling on value traded rather than share or contract count. A bar closes each time a fixed dollar amount changes hands.

The advantage appears over longer horizons and it is a real one. If a stock doubles, a fixed share-count bar now represents twice the economic activity it used to. A fixed dollar bar represents the same economic quantity throughout. The same argument applies to shares outstanding changing through issuance or buybacks.

For any study spanning years, or comparing instruments at different price levels, dollar bars are usually the most defensible choice.

Tick, imbalance and run bars

Two further variants are worth knowing.

Tick bars close after a fixed number of transactions. Simple, but vulnerable to trade fragmentation: one order broken into fifty prints generates fifty ticks, so the measure drifts as market structure changes.

Imbalance and run bars sample adaptively on order flow itself, closing when signed flow imbalance exceeds an expectation. These sample most densely exactly when flow becomes one-sided, which is when something is happening. They are more complex and more sensitive to parameterisation, and they are the natural extension of signed flow into the sampling decision itself.

The practical caveats

This is not an argument that every chart should be a dollar bar chart.

Sessions, events and human behaviour are genuinely time-based. The open, the close, an economic release and an expiry all happen at a clock time. Anything reasoning about time of day needs time.

Alignment across instruments is harder. Two symbols' volume bars close at different moments, so any cross-sectional comparison needs explicit handling.

Time bars are what everyone else uses, which matters for a reason beyond convention: levels that other participants watch are computed from time bars. A daily high is a daily high because the session defines it.

Alternative bars are unfamiliar to read. A chart where bar width varies with activity takes practice to interpret visually.

The sensible position is that these are different tools. Time bars for display, session logic, and anything referencing what other participants see. Volume or dollar bars for statistical work, model inputs, and any study whose conclusions depend on the distribution of the observations.

The transferable point

The reason this matters beyond charting is that a sampling scheme is an assumption, and assumptions that are never stated are never checked.

Choosing to sample by clock time is a modelling decision that was made for

you, by convention, and it is quietly wrong for most statistical purposes.

If a study's conclusion depends on the distribution of its observations, and almost every significance test does, then the sampling scheme is part of the method and belongs in the method description. Most published backtests do not mention it at all.