Telling Informed Flow From Uninformed Flow

Two traders each buy 1,000 shares of the same stock in the same minute, hitting the same market maker's offer at the same price. One is a pension fund rebalancing on a schedule set weeks earlier; the other has just learned, through legitimate research, that the company's next earnings will beat expectations by a wide margin. To the market maker, these two fills look identical at the moment they happen — same size, same price, same side. But one of them is a coin flip that costs the market maker nothing in expectation, and the other is a trade the market maker will lose money on as soon as the news becomes public. Sorting these two apart, even approximately, is the central problem of market making, and it is what the informed-versus-uninformed distinction is about.

What separates the two types

Uninformed (or "noise") flow trades for reasons unrelated to superior knowledge of where the price is going: portfolio rebalancing, index tracking, liquidity needs, hedging. On average, uninformed flow is priced fairly by the market maker — sometimes it's a good fill, sometimes a bad one, but there's no persistent directional edge against the market maker. Informed flow trades because the trader has a genuine edge on the security's near-term value — through research, faster processing of public information, or in the worst case actual non-public information. A market maker who cannot tell the two apart must price every trade as if it could be informed, which means quoting spreads wide enough to cover the expected losses from the informed share of flow. Glosten and Milgrom's 1985 model formalizes exactly this: the market maker sets bid and ask as conditional expectations of value given a trade at that price, and the spread that results is entirely a function of the fraction of flow believed to be informed — a market with zero informed traders would have zero spread.

A worked example

A market maker estimates that 20% of the order flow in a stock is informed, and that informed buyers only show up when the true value is $0.50 above the current mid, while uninformed buyers arrive regardless of value. The market maker's fair bid/ask must be set using Bayes' rule on the observed buy: the expected value conditional on seeing a buy order is a probability-weighted mix of the informed scenario (worth an extra $0.50) and the uninformed scenario (worth $0.00 extra), weighted by how much more likely a buy is to occur under each. If informed traders buy with certainty when informed and uninformed traders buy 50% of the time regardless, then out of every 100 orders, 20 are informed buys and 40 are uninformed buys (half of the 80% uninformed group), for 60 total buys. The probability a given buy is informed is $20/60 = 1/3$. The ask markup is $1/3 \times 0.50 = 0.17$, i.e. the ask should sit $0.17 above the pre-trade mid, purely to cover the expected $0.50 loss on the one-third of buys that are informed. If the informed share instead rises to 40% of flow, the same arithmetic (informed buys = 40, uninformed buys = 30, total buys = 70, informed fraction $= 40/70 \approx 0.57$) pushes the ask to roughly $0.29 above mid — nearly double the markup for exactly double the informed share, because the informed fraction of buys specifically rises faster than the raw informed fraction of all flow.

Where this gets used

• Spread setting: real market-making systems estimate a rolling proxy for informed-flow share (often via VPIN and Flow Toxicity or realized post-trade drift) and widen automatically when that estimate rises.
• Client tiering: dealers who internalize order flow classify counterparties by historical toxicity — see Scoring Counterparty Toxicity — and quote different effective spreads or reject flow from counterparties whose historical fills correlate with adverse price moves.
• Execution algo design on the buy side: a fund that trades in a way indistinguishable from noise (small, randomly timed clips) gets better fills than one that trades in a way that looks informed (large, urgent, directional), which is the whole rationale behind Detecting Metaorders In The Tape countermeasures like randomized order timing.

In interviews

If asked why bid-ask spreads exist beyond covering order-processing costs, the informed/uninformed distinction is the answer interviewers want: spread is compensation for adverse selection risk, priced via Bayes' rule on the base rate of informed flow, and it rises faster than linearly as that base rate rises — walk through the Bayesian update the way the worked example does rather than asserting the conclusion.