Bar Aggregation and Resampling Choices

Raw market data arrives as a stream of individual trades and quotes, each with its own timestamp. Almost no strategy works directly with that stream, instead, it gets grouped into "bars," each summarizing open, high, low, close, and volume over some chunk of the stream. The default choice, a bar every fixed clock interval like one minute, feels so natural that it's easy to forget it's a choice at all, and not a neutral one. During the first ten minutes after the open, hundreds of trades might occur; during a quiet lunchtime stretch, a one-minute bar might contain two trades or none. Fixed-time bars are built from data that arrives at a wildly uneven rate, which means the statistical properties of the bars, their volatility, their information content, vary through the day in ways that have nothing to do with the market and everything to do with the clock.

The alternatives

Volume bars close a bar after a fixed amount of shares or contracts have traded, so a bar during the busy open might last ten seconds while a bar during a quiet stretch might last twenty minutes, but each bar represents roughly the same amount of trading activity. Dollar bars do the same thing using traded notional value, which is more stable than volume when a security's price changes a lot over the sample period. Tick bars close after a fixed number of individual trades. All three adapt to the market's own pace rather than the clock's, and research comparing them has generally found that volume- and dollar-based bars produce returns closer to a normal distribution and with more stable variance than time bars, a genuinely useful property if the strategy or its statistical tests assume anything about the shape of the return distribution.

A concrete case

Consider building a mean-reversion signal on one-minute bars for a stock around an earnings announcement. In the minute right after the release, thousands of trades might occur and the price can move several percent; an hour later, the same one-minute window might see a handful of trades and barely move. A signal tuned on the quiet-period bars will be systematically miscalibrated for the volatile period, because a "one-minute move" means something completely different in the two regimes. Rebuilding the same data as volume bars instead means each bar reflects a comparable amount of market activity throughout the day, so the signal sees a more consistent unit of "how much happened" regardless of when it happened.

What this means in practice

The choice of bar type is a modeling decision, not a data-formatting detail, and it should be made deliberately and documented alongside the rest of the strategy's assumptions. Time bars remain the right choice when a strategy is inherently about calendar time, end-of-day rebalancing, or anything tied to scheduled events, but for signals meant to capture market microstructure or momentum, activity-based bars are usually a better match for how information actually arrives.