Bar Construction Methods

The most familiar way to summarize a day of trading is a time bar: open, high, low, close for every fixed interval — every minute, every five minutes. It's the default in almost every charting tool, and it's also, statistically, a somewhat arbitrary choice: a minute of trading during a market open, with hundreds of trades crashing through, and a minute during a lunchtime lull, with two or three, get compressed into the same-sized bucket even though wildly different amounts of information happened in each. That mismatch degrades the statistical properties — returns computed from time bars are less close to independent and identically distributed than practitioners often assume, which matters if a model downstream is implicitly relying on that assumption.

Alternative bar types sample based on activity instead of the clock. A tick bar closes after a fixed number of trades (say, every 500 trades) regardless of how long that took. A volume bar closes after a fixed number of shares/contracts have traded. A dollar bar closes after a fixed dollar value has traded — arguably the most robust of the three, because it self-adjusts for a stock's price changing over time (a $1,000,000 dollar bar means the same "amount of trading" whether the stock is at $50 or $500, while a fixed share-count volume bar doesn't). All three produce more bars during high-activity periods and fewer during quiet ones, which tends to sample the market more evenly in terms of information content rather than clock time, and empirically produces returns with statistical properties (closer to normal, less serial correlation) that are more favorable for many modeling techniques than time bars.

Worked example: comparing bar counts across a session

A liquid stock trades $2,000,000 in total dollar volume during a 30-minute burst around an earnings announcement, and $200,000 total during a quiet 30-minute stretch in the early afternoon. Using 1-minute time bars, both periods produce exactly 30 bars each, despite the burst period having ten times the trading activity of the quiet one — most of the real information is crammed into the burst's 30 bars, and the quiet period's 30 bars are comparatively empty and noisy. Using $100,000 dollar bars instead, the burst period produces $2{,}000{,}000 / 100{,}000 = 20$ bars and the quiet period produces $200{,}000/100{,}000 = 2$ bars — ten times as many bars where ten times as much actually happened, giving a downstream model roughly uniform "information per bar" rather than uniform "clock time per bar."

What this means in practice

Bar choice matters most when bars feed a statistical or machine-learning model, where the closer-to-i.i.d. behavior of activity-based bars can materially improve model quality; for simple visual charting, time bars remain the practical default because traders are used to reading them. Dollar bars are generally preferred over volume or tick bars for cross-asset or long-horizon work because they're naturally robust to price-level changes and stock splits, which volume and tick bars are not.