Micro-Price vs Mid-Price: The Fair Value Your Chart Hides
The mid-price is an average of two prices that ignores the sizes behind them — and the sizes are the market. The micro-price corrects that in one multiplication, and a one-tick classifier built on queue imbalance alone reaches an out-of-sample AUC of 0.75–0.80 on large-tick stocks: a real, mostly mechanical effect that still is not an edge. The published numbers with their real names, and the 16 ms plumbing behind the indicator we just shipped.
Every platform shows you the mid-price. Almost none shows you where fair value actually sits — which is strange, because the correction has been published for years, costs one multiplication per book update, and the raw material is already in the level-2 feed you are paying for. This is the field note: the published numbers with their real names, why the effect is closer to physics than to alpha, and the plumbing that let us ship it as a chart overlay for free.
The numbers
- 0.75–0.80 — out-of-sample AUC of a classifier predicting the direction of the next mid-price move from queue imbalance alone, across the large-tick stocks in Gould & Bonart (2016). On small-tick names it drops to around 0.6 — a moderate improvement over chance, much weaker.
- ~80–90% — how often the next mid-price move goes with the heavy side when the imbalance is pinned near ±1, per the same paper’s fitted curves. The effect is real; what it is worth is the next number.
- 1 tick — the horizon at which all of that predictability lives.
- 16 ms — the cadence of the coalesced depth stream our charts already consume; the micro-price and QI recompute on every update.
- 0 — extra data subscriptions, trade classifiers or fitted models the whole thing requires.
What the AUC actually says
The signal-sellers’ version of this research is “the book predicts the next move with 75% accuracy”. That is not what the paper measures. AUC is a ranking statistic: 0.75–0.80 means the model scores a randomly chosen up-move above a randomly chosen down-move 75–80% of the time. It is strong evidence the queues carry information — and it is not a hit rate you could trade off.
And most of that information is mechanical. When forty contracts rest on the bid and ten on the ask, the thin queue is the one consumed first by construction — predicting it is closer to physics than to alpha. To monetize one tick you need front-of-queue position and microsecond latency, and that business is spoken for. So we ship the numbers the same way we ship OFI: as a state read, never as an entry arrow.
Why the mid lies, in one sentence
The mid-price averages two prices as if the sizes behind them did not exist; the micro-price weights the same two prices by the opposite queue’s share, so fair value slides toward the thin side — and it slides before anything trades, because the queues shift before the print does. The full walkthrough, with a worked book and every setting, is in the guide: how the Microstructure indicator works.
The plumbing that makes it free
The academic pipeline for this kind of study is heavy: reconstruct the limit order book event by event from an exchange dump, sign the flow, fit and validate the classifier. The practitioner’s version, it turns out, is almost nothing. Our charts already hold a live level-2 book — the same stream that feeds the heatmap and the DOM ladder — coalesced to a 16 ms cadence over Rithmic. On each update, the micro-price is one multiplication, QI is one division, and ten-level OFI is a handful of subtractions. No trade classifier (a CME feed arrives aggressor-signed, as the OFI post details), no model to fit, no additional subscription — the marginal cost of the entire indicator is arithmetic on data you already pay for. That asymmetry, more than any single number, is the reason it exists: the research is a decade old, the inputs are in every serious futures feed, and yet almost no retail platform draws it.
The config we ship
- QI — best-level only, the formulation the published result is built on.
- Micro-price trail — 180 seconds of history on the chart.
- OFI — 10 levels, depth-normalized, EMA-smoothed over 20 book events.
- Gauge dead zone — ±0.15: a balanced book reads as no signal, on purpose.
Where it earns its place
Levels first, book second. The levels come from wherever your process gets them — profile, absorption, GEX. When price arrives, the gauge answers the only question the book can answer honestly: is it leaning with you or against you, right now? Both numbers leaning your way while price holds is the book on your side. Both against you, and the wall you are leaning on is being quietly dismantled — information the mid-price never carried.
Frequently asked questions
- An AUC of 0.75–0.80 sounds high — why is that not a strategy?
- Two reasons. First, AUC is a ranking statistic, not a hit rate: it says the classifier scores a randomly chosen up-move above a randomly chosen down-move 75–80% of the time, which is weaker than "calls 75% of moves correctly". Second, the effect lives at the one-tick horizon and is mostly mechanical — the shorter queue is consumed first. Monetizing one tick requires front-of-queue position and latency measured in microseconds, and that business is spoken for. As a chart read at a level you already care about, the number is useful; as a standalone signal, it is a way to donate spread.
- Is the micro-price the same as the Stoikov micro-price?
- What we draw is the imbalance-weighted mid — the practical first-order form. Stoikov’s full 2018 estimator refines it with a Markov-chain adjustment for how imbalance itself evolves. The weighted mid captures most of the behavior, computes in one multiplication per book update, and is what most practitioners mean by the term day to day.
- Does this replace watching the DOM?
- No — it compresses one specific aspect of it. The DOM shows you the full ladder: walls, icebergs, pulled orders at every level. The micro-price and QI condense just the best-level queue balance into a line and a number you can keep in peripheral vision while your eyes stay on the chart. Traders who read the DOM well will recognize the information; the indicator makes it visible without the ladder.
Order Flow on a Prop-Firm Account: What Actually Works on Apex, Topstep and 4PropTrader
The rules, the feeds and the drawdown maths nobody writes about because they trade their own money.
v0.7.4: Wiring the Full Order Book (MBO), and a Strategy That Became an Indicator
From ~15 visible levels to the whole book: the MBO rebuild, a heatmap that finally shows two sides, and signals with zero order routing.