OFI vs CVD: What the Tape Cannot See, Measured

Every order-flow trader knows CVD. Almost nobody outside of academic desks uses OFI — which is strange, because the published gap between them is not small. This post is the measured version of the argument, with the numbers, where they come from, and the exact configuration we ship.

The numbers

• ~87% — share of contemporaneous price variance explained by integrated ten-level OFI (Cont, Cucuringu & Zhang, 2023).
• ~71% — same measure, best level only (the original Cont–Kukanov– Stoikov formulation).
• ~33% — signed trade volume, the raw material of CVD.
• Negative — the out-of-sample R² when lagged OFI is used to forecast the next minute. Not small. Negative.

Read those four lines together and you get the honest story: the order book explains price far better than the tape does, and neither one predicts. Any product that sells OFI as a forecasting signal is selling the fourth bullet upside down.

What the tape physically cannot see

CVD only increments when a trade prints. But watch a DOM during a real move: bids stack three levels deep before the breakout trade ever fires; a wall of asks quietly evaporates and price drifts up through the vacuum without a single aggressive buy. None of that touches CVD. All of it is OFI’s raw material — adds, pulls and queue depletion, level by level. That is the mechanical reason behind the 87-versus-33 gap: most of the price formation happens in the book, between the prints.

The plumbing detail that makes it free on futures

The academic pipeline needs a trade classifier (Lee-Ready, BVC) to sign the tape, because equity feeds do not carry the aggressor. A CME feed over Rithmic does — every trade arrives already signed, and the L2 book updates arrive on the same session. Which means the entire OFI computation is a subtraction per level per update: no classifier, no model, no extra data subscription. It was, frankly, sitting in the feed all along.

The configuration we ship

In Senzoukria, the Microstructure (quant) indicator computes three things from the same book stream: the micro-price (imbalance-weighted fair price inside the spread), the best-level queue imbalance (QI), and integrated OFI. The defaults, which you can copy or change:

• OFI levels: 10 — the integrated depth from the research; 1 gives you the classic best-level formulation.
• Normalization: always on — OFI is divided by a slow average of depth per level. A raw OFI is not comparable across hours, let alone instruments.
• Smoothing: 20 book events — the displayed value is an EMA; per-event OFI is flicker.

How we actually use it: not as an entry signal — the fourth bullet above forbids it — but as the book’s opinion at a level chosen for other reasons. Price into an absorption zone with QI pinned positive and OFI climbing is a different trade from the same level with the book leaking the other way. The full mechanics are in the companion guide: OFI explained.