VWAP: what it is, and whether it works

735 of 3,344 out-of-sample tests beat simply buying and holding the same asset — 22.0%. On the other 2,609 it did not. That is above the 20.1% rate across all 382 indicators we test (one pooled rate over all 660,005 tests we have run, not a mean of the per-indicator rates).

One test is one variant on one asset on one timeframe. We tested 2 variants of VWAP, so an asset that carries several of them contributes several tests — this is the family's record across everything we ran, not a count of assets. There are only two, and they differ: 19.7% and 24.6%. 22.0% is the pooled rate over both, not a midpoint and not a property either one has on its own — the table below gives each separately. "Beat" means a higher return than holding that same asset over that same window. Measured out-of-sample, on data the setup was not chosen on.

VWAP is a running total of price times volume divided by a running total of volume, restarted at the beginning of each session. The price used for each bar is conventionally the typical price, (high + low + close) / 3. Because every bar is weighted by the volume that traded in it, a large block moves VWAP far more than a quiet bar at the same price, which is what separates it from a simple moving average of the same length.

How it's read. It is used as a fair-value reference rather than a signal generator. Price above VWAP means the average participant in the session so far is holding a profit; below it, a loss. Institutional desks are frequently measured against VWAP as an execution benchmark, which is a large part of why the level attracts activity at all.

Where it struggles by design. It resets every session, so it says nothing about trend beyond the current day and cannot be read across days the way a moving average can. Early in a session it is computed from very little volume and is correspondingly unstable, and on thin instruments it can be dragged a long way by a single print.

Origin: Originated as an institutional execution benchmark in the mid-1980s (first benchmarked executions 1984; formalised in the Journal of Finance in 1988); wide use followed with algorithmic execution.